diff --git a/README.md b/README.md index c315ece..b2315e4 100644 --- a/README.md +++ b/README.md @@ -45,6 +45,8 @@ The VIPeR download page is `/datasets/viper/`. Its ZIP is the exact historical v The essay's performance scatter plot uses `assets/rank-one-scatter.json`, which records reported results, evaluation settings, dates, and paper sources. Run `python3 build_rank_one_chart.py` with Matplotlib installed to regenerate `assets/rank-one-scatter.png`. The plot distinguishes training settings and is not a controlled leaderboard. +The essay's informal Haiku/Opus experiment is documented separately in `assets/foundation-model-evaluation.json`, including prompts, aggregate counts, and evaluation limits. It is a single-split observation, not an additional literature-chart milestone or a five-repeat benchmark estimate. + The `last_updated` field in `site/resume.json` is also the source of truth for the "Last updated" text shown on the website, the PDF resume and the generated `llms.txt` summary. When changing profile content or rebuilding artifacts, update this date in the JSON. ## LLM-readable summary diff --git a/site/writing/twenty-years-of-person-re-identification/assets/foundation-model-evaluation.json b/site/writing/twenty-years-of-person-re-identification/assets/foundation-model-evaluation.json new file mode 100644 index 0000000..4ecb14e --- /dev/null +++ b/site/writing/twenty-years-of-person-re-identification/assets/foundation-model-evaluation.json @@ -0,0 +1,84 @@ +{ + "title": "Exploratory prompted-model evaluation on VIPeR", + "version": 1, + "completed_date": "2026-10-14", + "date_timezone": "America/Los_Angeles", + "scope": "Clothing and scene appearance retrieval, followed by listwise reranking. This is an informal baseline, not a controlled leaderboard or a human-ceiling study.", + "dataset": { + "name": "VIPeR v1.0", + "identities": 632, + "archive_sha256": "857e49928a0c74d2da6fbeef372ea2e42b2bce357f87307d74c6601f1a97f267", + "input_image_width": 96, + "input_image_height": 256, + "image_preprocessing": "Bilinear resize from the original 48 by 128 pixel crops." + }, + "evaluation": { + "repetitions": 1, + "test_identity_count": 316, + "gallery_size": 316, + "directions": ["camera A to camera B", "camera B to camera A"], + "total_queries": 632, + "other_identity_half": "Not evaluated and not used for weight training.", + "partition_seed": 20261013, + "test_half": 1, + "ranking_seed": 20261014, + "original_prompt_development_identity_count": 50, + "development_identities_in_this_test_half": 13, + "prompt_held_out_identity_count": 303, + "prompt_held_out_query_count": 606, + "tie_handling": "Fixed seeded ordering, shared by the retrieval and reranked comparison.", + "invalid_output_handling": "A malformed ranking counts as unsuccessful at every CMC rank; no ranking is guessed or silently retried." + }, + "retrieval": { + "model": "claude-haiku-5-5", + "effort": "low", + "thinking": "adaptive", + "max_output_tokens": 1024, + "system_prompt": "You score visual appearance compatibility for a pedestrian retrieval research evaluation.", + "prompt": "Compare these pedestrian crops as a possible cross-camera match. Use non-facial appearance: clothing, patterns, accessories, carried objects, body outline, and scene context. Allow changes in viewpoint and lighting. Score the overall evidence for a match from 0 to 1000: 0 is clearly incompatible and 1000 is strongly compatible. Use only the two attached images. Do not use tools or inspect files. Return only one JSON object: {\"score\": }.", + "score_matrix_shape": [632, 632], + "reverse_direction": "Ranks the transpose of the existing A-first/B-second pair scores, rather than independently rescoring reversed prompts." + }, + "reranking": { + "model": "claude-opus-5-5", + "effort": "high", + "max_output_tokens": 8192, + "shortlist_size": 30, + "system_prompt": "You rank clothing and scene appearance for a cross-camera retrieval evaluation.", + "prompt": "Rank the candidate crops by their visual compatibility with the probe crop. Base your ranking only on clothing, patterns, accessories, carried objects, outfit outline, and scene context. Allow changes in viewpoint and lighting. Give highest priority to putting the most compatible candidate first, then order the remaining candidates from best to worst. Use only the attached images; do not use tools or inspect files. Return only one JSON array of the supplied candidate labels in ranked order. Include every candidate label exactly once. Do not output scores, explanations, Markdown, or any other text. The labels and their presentation order carry no information about compatibility.", + "presentation": "One probe and 30 anonymously labeled, shuffled candidates per fresh session, with no tools or conversation history. The correct counterpart is not inserted if absent from the shortlist.", + "ranking_beyond_shortlist": "Preserves the original Haiku order after the reranked top 30.", + "completed_queries": 632, + "valid_rankings": 629, + "invalid_rankings": 3, + "refusals": 0 + }, + "results": { + "prompt_held_out": { + "queries": 606, + "haiku_rank_1_correct": 296, + "opus_reranked_rank_1_correct": 355, + "haiku_rank_5_correct": 412, + "opus_reranked_rank_5_correct": 452, + "haiku_rank_10_correct": 466, + "opus_reranked_rank_10_correct": 484, + "rank_1_corrections": 94, + "rank_1_regressions": 35 + }, + "all_test_queries": { + "queries": 632, + "haiku_rank_1_correct": 310, + "opus_reranked_rank_1_correct": 368, + "rank_1_corrections": 96, + "rank_1_regressions": 38 + } + }, + "limitations": [ + "Only one frozen identity split was evaluated with Opus, not five repetitions.", + "Prompt selection used 50 development identities; the headline subset excludes those identities, but earlier failure diagnostics inspected some other evaluation images. This is not a pristine blind test.", + "No task-specific weight training was performed, but the models' original training data and any prior exposure to VIPeR are unknown.", + "Published specialist results use different source training datasets, split averaging, and sometimes different fitting or reranking procedures.", + "The prompts deliberately score clothing and scene appearance; the experiment does not establish an unrestricted capability ceiling for either model.", + "This file records aggregate measurements and the protocol, not a complete reproduction bundle." + ] +} diff --git a/site/writing/twenty-years-of-person-re-identification/assets/viper-paired-sample-authors.png b/site/writing/twenty-years-of-person-re-identification/assets/viper-paired-sample-authors.png new file mode 100644 index 0000000..8c8c133 Binary files /dev/null and b/site/writing/twenty-years-of-person-re-identification/assets/viper-paired-sample-authors.png differ diff --git a/site/writing/twenty-years-of-person-re-identification/index.html b/site/writing/twenty-years-of-person-re-identification/index.html index 525bf12..ef2db3e 100644 --- a/site/writing/twenty-years-of-person-re-identification/index.html +++ b/site/writing/twenty-years-of-person-re-identification/index.html @@ -9,7 +9,7 @@ - + Reflections on 20 Years of Person Re-ID - Douglas Gray @@ -21,7 +21,7 @@ "author": {"@type": "Person", "name": "Douglas Gray", "url": "https://de3ug.github.io/"}, "dateCreated": "2026-10-12", "datePublished": "2026-10-13", - "dateModified": "2026-10-13", + "dateModified": "2026-10-17", "url": "https://de3ug.github.io/writing/twenty-years-of-person-re-identification/" } @@ -32,52 +32,54 @@
- +

Reflections on 20 Years of Person Re-ID

Measuring the problem, watching the field grow, and reconsidering the purpose of the work.

VIPeR dataset: download, details, and citation

-

In 2006, I read Gheissari and colleagues' paper on recognizing people across cameras. The accuracy looked impressive, but the evaluation used just 44 people, seen largely from the front.1 That was a much easier problem than the one I recognized from my undergraduate work.

+

Twenty years ago, I started building VIPeR to test whether a computer could recognize the same person across different cameras. It's a small dataset by today's standards: two images each of 632 people. But it became a widely used benchmark, and those tiny images still give today's AI models trouble. Revisiting it has reminded me how much progress starts with deciding what to measure, collecting the right data, and being willing to get a disappointing result. That's a lesson I've carried through much of my career. It's also brought me back to a less comfortable part of the story. VIPeR gave me a promising start in surveillance research, just as I was becoming unsure that making people easier to track was what I wanted to do. What would happen when we got good at this?

-

I had been working on crowd counting: estimating how many people were in a scene, perhaps 50 or 100 of them moving through it, often seen from an awkward camera angle. People faced different directions, overlapped, and appeared at different scales. A collection of mostly frontal images did not represent that variety.

+

The appearance problem between cameras

-

The issue was not whether a recognition method could work in the setting that paper studied. It was what the results told us about finding someone from an arbitrary camera angle, with different lighting, a different pose, and hundreds of plausible matches. How could we make progress on the real problem without measuring it properly, even if a realistic test seemed impossibly difficult?

+

Multi-camera surveillance systems were becoming an active research area. Tracking within one camera was already difficult, but expanding coverage introduced another question: when a person appeared in a second camera, was it someone the system had seen before?

-

That question led me to start collecting the Viewpoint Invariant Pedestrian Recognition dataset, VIPeR. The first paper appeared in 2007, but the work began in 2006. Twenty years later, the benchmark has been cited thousands of times and helped researchers develop methods well beyond anything we could do then. It also made surveillance research a promising career path for me. I had an early foothold in a growing field, a dataset others were adopting, and a clear supply of difficult problems to work on. Continuing would have made sense. I eventually chose a different direction.

+

Overlapping fields of view could provide geometric constraints. Nearby cameras could provide timing constraints: someone who left one view might enter another a few seconds later. Those cues were useful, but could become weak or unavailable when cameras were farther apart, their views didn't overlap, or the task was to search old footage without knowing when the person would appear.

-

The appearance problem between cameras

+

Appearance was another source of evidence, and it was the one a human would naturally use. A person looking through video could compare clothing, shape, texture, and what someone was carrying. We needed a model that could make those comparisons despite changes in viewpoint and illumination, and a reliable way to test it across that variation.3

-

Multi-camera surveillance systems were becoming an active research area. Tracking within one camera was already difficult, but expanding coverage introduced another question: when a person appeared in a second camera, was it someone the system had seen before?

+

In 2006, I read Gheissari and colleagues' paper on recognizing people across cameras. The accuracy looked impressive, but the evaluation used just 44 people, seen largely from the front.1 That was a much easier problem than the one I knew from undergraduate research on crowd counting.

+ +

The issue wasn't whether a recognition method could work in the setting that paper studied. It was what the results told us about finding someone from an arbitrary camera angle, with different lighting, a different pose, and hundreds of plausible matches. How could we make progress on the real problem without measuring it properly, even if a realistic test seemed impossibly difficult? That question led me to start collecting the Viewpoint Invariant Pedestrian Recognition dataset, VIPeR.

+ +

Building VIPeR

-

Overlapping fields of view could provide geometric constraints. Nearby cameras could provide timing constraints: someone who left one view might enter another a few seconds later. Those cues were useful, but could become weak or unavailable when cameras were farther apart, their views did not overlap, or the task was to search old footage without knowing when the person would appear.

+

In 2003, I took Professor Hai Tao's undergraduate multimedia class. There were only three students. The next quarter, at his invitation, I was taking graduate computer vision, and before I graduated I was doing research with his Ph.D. students. He recognized my interest and ability and gave me a place in the research community. I stayed on for the Ph.D. with him as my adviser, and I owe him my start in the field.

-

Appearance was another source of evidence, and it was the one a human would naturally use. A person looking through video could compare clothing, shape, texture, and what someone was carrying. We needed a model that could make those comparisons despite changes in viewpoint and illumination. Simple color histograms were common, but they lost spatial information and could confuse people in similar clothing. Templates preserved layout but changed dramatically with pose. Existing models offered pieces of a solution; we lacked a reliable way to compare them on the full range of variation.3

+

My first real research project was crowd counting: estimating how many people were in a scene, sometimes 50 or 100 of them moving through it. People overlapped, faced different directions, and appeared at different scales, often from an awkward camera angle. We were trying to count them rather than recognize them, but those images shaped my sense of what a realistic recognition problem would have to include.

-

My earlier crowd-counting work had exposed me to people from all angles and camera orientations. A benchmark for recognition needed to include that variation and enough different identities to make matching meaningful. High accuracy on a small, easy evaluation could not tell us which appearance models would work in a camera network.

+

Some of the other work in the lab was adapting color correlograms for object tracking.31 In content-based image retrieval, or CBIR, we were trying to represent appearance with hand-designed color and texture statistics. A histogram could tell us how much blue or gray was present, but not how those colors were arranged. Correlograms added information about how often colors occurred at different distances from one another. Looking back, these were primitive descriptions of a person. At the time, finding a way to preserve even a little more structure was useful progress.

-

Building VIPeR

+

That dependence on color extended to the cameras themselves. When we set out to collect VIPeR, we used two video cameras Hai had purchased. I still remember his enthusiasm for their three-CCD design, with separate sensors for red, green, and blue. Today we take good color video for granted; then, capturing it well was part of getting the appearance model to work.

-

My adviser had purchased two cameras and made a point of telling us they were three-CCD cameras, with separate sensors for red, green, and blue. It is easy to take good color video for granted now that we all carry cameras in our phones. Back then, the camera hardware and its low-level details were a noticeable part of the experiment, especially when so much of an appearance model depended on color.

+

We hired an undergraduate to help capture footage, with both cameras running at the same time. Once we had the recordings, the challenge was finding the same person in both views. I built a tool to draw bounding boxes and record the correspondences, but each match still had to be found by hand. Building the benchmark meant doing, over and over, the very task we wanted a computer to help with.

-

We ran two cameras at the same time and hired an undergraduate to help capture footage. I built an annotation tool to draw bounding boxes around people and establish correspondences between the views. The tedious part was finding the same person manually in both recordings. That was precisely the task we wanted the technology to help with.

+

That work helped shape how I've approached research ever since. Most of the important projects in my career have begun with the same questions: what data do we need to collect, how should we annotate it, and would a better result on this evaluation mean progress on the problem we actually care about?

-

The final dataset contained two images of each of 632 identities: 1,264 images in total, each normalized to 48 pixels wide by 128 pixels tall. Most pairs differed in viewpoint by at least 90 degrees. The range of identities and the variation between the paired views were what made it useful.2

+

The final dataset contained two images of each of 632 identities: 1,264 images in total, each normalized to 48 pixels wide by 128 pixels tall. Most pairs differed in viewpoint by at least 90 degrees. The range of identities and the variation between the paired views were what made it useful.2 My coauthor Shane Brennan (ID 246) and I (ID 050) are in there too.

-
- Twelve VIPeR identity pairs showing the same pedestrians from different camera viewpoints +
+ Twelve paired VIPeR identities, labeled by dataset ID, with one image from each camera
Sample identity pairs from VIPeR. Each pair contains one image from each camera. Download the dataset and read its documentation.
-

The evaluation protocol was another contribution of the PETS paper. Ranking was not new; the 2006 paper that prompted this work also reported ranked matches. Other work treated recognition as a binary same-person/different-person decision and reported ROC curves. We argued for evaluating the actual search task: given a probe image and a gallery of candidates, where does the correct person appear? A low false-positive rate on individual pairs does not directly tell you whether the right match will be near the top of a large gallery.

+

The evaluation protocol was another contribution of the PETS paper. We wanted to measure appearance matching separately from the geometry and timing constraints of a complete tracking system. A good end-to-end result could depend on overlapping camera views or predictable travel times without revealing how well the appearance model worked on its own. VIPeR let us compare those models without relying on either cue.

-

We used the cumulative matching characteristic, or CMC: how often does the correct person appear within the first k results? We did not invent CMC, but making ranking and gallery size explicit was an important part of the evaluation we proposed. It connected appearance matching to the task of helping an operator find someone, rather than reducing the problem to classifying isolated pairs. We also showed how to relate the curve to simplified tracking and reacquisition scenarios.2

- -

To compare methods consistently, we prescribed 5x2 cross-validation: five randomized two-fold rounds, with randomized image order and mirroring, and averaging both gallery/probe directions. This let us study the appearance model separately from the geometric and timing assumptions of a complete tracking system.

+

We evaluated appearance matching as a search task: given an image of a person and a gallery from another camera, where did the correct match appear? We used the cumulative matching characteristic, or CMC: how often was it among the first k results? Making ranking and gallery size explicit connected the evaluation to the task of narrowing a search for a human operator.2

Too difficult to publish?

-

Early versions were rejected from CVPR and ICCV before the work found a home at the PETS workshop in 2007. My recollection is that reviewers cited the low accuracy as one of the main flaws: they did not believe the method worked. From my perspective, that difficulty was evidence that we had finally included important parts of the problem. We needed to measure the gap before we could close it.

+

Early versions were rejected from CVPR and ICCV before the work found a home at the PETS workshop in 2007. As I remember it, reviewers cited the low accuracy as one of the main flaws: they didn't believe the method worked. To me, that difficulty meant we'd finally included important parts of the problem. We needed to measure the gap before we could close it.

The best rank-one baseline in the 2007 PETS paper, a color correlogram with manually localized body regions, reached roughly 5 percent, as read from its CMC plot.2 Our Ensemble of Localized Features, published at ECCV in 2008, reached about 12 percent. ELF used AdaBoost to select useful local color and texture features and combine their comparisons into a learned similarity function.3 It was progress, but clearly left most of the problem unsolved.

@@ -85,112 +87,63 @@

Too difficult to publish?

We quantified that idea with an expected-search-time model. Assuming an operator could review one image per second without making mistakes, the expected search took 158 seconds with an unsorted gallery and 28.9 seconds with ELF: an 81.7 percent reduction. These were modeled search times, not timings from a user study, but they made the intended benefit explicit.3

-

What changed as the field grew

- -

VIPeR gave researchers a shared test of appearance matching across viewpoints. A few influential methods show how the field progressed from designing better features to learning how to compare them.

- -

In 2010, Farenzena and colleagues' Symmetry-Driven Accumulation of Local Features, or SDALF, organized color and texture around body regions and their symmetry. Weighted color histograms, stable color regions, and distinctive repeated patches retained more of a person's visual structure while allowing changes in pose and viewpoint. The paper reported roughly 20 percent rank-one accuracy on VIPeR, compared with about 12 percent for ELF.4

- -

Köstinger and colleagues' 2012 Keep It Simple and Straightforward MEtric, or KISSME, made metric learning computationally simple. Given pairs labeled as the same or different people, it modeled their feature differences and derived a distance metric from a likelihood-ratio test. Some variation is normal between two views of one person; other variation distinguishes identities. KISSME supplied an efficient, reusable way to learn that distinction. Its VIPeR result was 19.6 percent at rank one, averaged over 100 runs.5

- -

The next few years brought more careful treatment of distinctive details and misalignment. Zhao, Ouyang, and Wang's 2013 salience methods emphasized patches that distinguished one person from others, rather than treating every region equally. Their unsupervised eSDC feature combination reached 26.74 percent rank one; supervised SalMatch reached 30.16 percent.1617 In 2014, Yang and colleagues' Salient Color Names work combined color-name distributions and color histograms across color spaces, reaching 37.8 percent with its fused representation.18

+

What changed as the field grew

-

Liao and colleagues' 2015 work paired Local Maximal Occurrence features with Cross-view Quadratic Discriminant Analysis. LOMO pooled local color and texture occurrences across horizontal positions, allowing an informative detail to move as the viewpoint changed. Illumination handling made the representation more stable; XQDA learned a discriminative subspace and comparison metric together. The combination reached 40 percent rank one on VIPeR and became a strong baseline for later work.6

+

Looking back, what stands out is how much researchers kept finding in these tiny crops. SDALF and LOMO/XQDA made better use of body layout, color, and texture, while metric learning helped decide which differences mattered.456 Neural methods then learned the features as well as the comparison.8 A shared benchmark let us see what each approach contributed, and what it still missed.

-

Progress also came from combining comparison models and using the gallery itself. Paisitkriangkrai and colleagues' 2015 metric ensembles optimized ranking across several features and metrics, reaching 45.89 percent. García and colleagues' DCIA used the context of visually similar gallery images to refine an initial ranking, reaching 63.92 percent with KCCA.1920 That distinction matters: re-ranking a gallery has access to information beyond comparing one probe with one candidate.

- -

Matsukawa and colleagues' 2016 Gaussian of Gaussian descriptor modeled local appearance distributions using both means and covariances, then described how those distributions varied across a region. Combined with XQDA and multiple color spaces, it reached about 50 percent rank one on VIPeR.7 Rich hand-crafted representations still competed well on a dataset with only 316 training identities per fold.

- -

That small training set was also the motivation for Zhang, Xiang, and Gong's 2016 discriminative null-space method. It sought a projection that collapsed same-identity training examples together while preserving separation between identities. Its fused-feature result reached 51.17 percent rank one.21

- -

Meanwhile, DeepReID and subsequent neural methods learned representations and matching jointly.8 Larger datasets made that approach increasingly effective. VIPeR continued to appear in tables, often as a transfer target, but no longer provided enough training data to represent the field's main ambitions. The 2016 survey Person Re-Identification: Past, Present and Future offers a useful account of this transition.9

- -

Later work asked a different question: could a model trained on other datasets recognize people in VIPeR without training on VIPeR itself? DIMN and DualNorm were already testing that idea in 2019, reporting 51.23 and 59.4 percent respectively.2829 Meanwhile, an expanded OSNet study reached 68 percent after fine-tuning on VIPeR, illustrating why the training setting matters.30

- -

SNR's style normalization and restitution reported 52.9 percent in 2020; RaMoE's relevance-aware mixture of experts reached 56.6 percent in 2021 under the same five-source setting. ACL reached 66.4 percent in 2022 using a different, four-source pool, and ReNorm reached 69.1 percent in 2024.22232425 In 2025, CLIP-FGDI adapted a pretrained vision-language model using learned prompts and reported 76.0 percent in its main VIPeR comparison.26 These results belong to the benchmark's history, but they are not a continuation of one unchanged experiment.

+

The more interesting change was in the role of the dataset itself. VIPeR offered little data for learning a representation from scratch, but a useful test of one trained elsewhere: could it handle a camera pair it had never seen?92829 CLIP-FGDI, for example, adapted a pretrained vision-language model and reported 76 percent rank-one accuracy in 2025.26 The chart records both better matching methods and the growing amount of training data behind them.

Scatter plot of reported VIPeR rank-one accuracy from 2007 to 2026, with selected methods labeled and NK3ML's 99.8 percent marked as a reported claim.
Reported VIPeR rank-one accuracy over time. Training data and evaluation settings vary across methods. Data and paper references.
-
- Values for labeled methods - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Year and methodRank one
Historical benchmark methods
2007 PETS baseline~5%
2008 ELF~12%
2010 SDALF~20%
2012 KISSME19.6%
2013 eSDC26.74%
2013 SalMatch30.16%
2014 SCNCD37.8%
2015 LOMO + XQDA40.0%
2015 Metric ensembles45.89%
2015 KCCA + DCIA (re-ranking)63.92%
2016 GOG + XQDA49.7%
2016 NFST fusion51.17%
2019 OSNet + VIPeR fine-tuning68.0%
Unseen-target domain generalization
2019 DIMN51.23%
2019 DualNorm (ResNet50)59.4%
2020 SNR52.9%
2021 RaMoE56.6%
2022 ACL66.4%
2024 ReNorm69.1%
2025 CLIP-FGDI76.0%
Other reported results
2018 NK3ML (reported claim)99.8%
2026 RND73.8%
-
- -
+

Use of the benchmark grew through much of the 2010s. Annual citations to the PETS and ELF papers peaked in 2017 and 2018. By October 2026, Google Scholar listed 1,405 and 2,016 citations respectively. New papers still cite them, but the pace has slowed as the field has moved to larger and more varied datasets.

+ +
Google Scholar citation histories: the PETS paper peaks in 2017 and the ECCV paper in 2018
Annual citations to the two foundational papers, from their Google Scholar histories observed on October 12, 2026. The papers are shown separately; 2026 is incomplete. Chart data and sources.12
-

The citation histories reflect that change. The PETS paper's annual citations peaked at 179 in 2017; the ECCV paper's peaked at 213 in 2018. Google Scholar now lists 1,405 and 2,016 citations respectively. New papers still cite them, but the pace has slowed as the field has moved to larger and more varied benchmarks.

-

Outgrowing the benchmark

-

Market-1501 made the limitations of the older benchmarks explicit: small identity sets, few cameras, hand-drawn boxes, and too little variation to represent deployment. It introduced more than 32,000 detector-produced boxes for 1,501 identities across as many as six cameras.10 MARS extended the task to video tracklets.11 Later datasets added occlusion, infrared imagery, larger galleries, domain shift, and people who were absent from the gallery altogether.

+

In 2015, Market-1501 made the limitations of the older benchmarks explicit: small identity sets, few cameras, hand-drawn boxes, and too little variation to represent deployment. It introduced more than 32,000 detector-produced boxes for 1,501 identities across as many as six cameras.10 MARS extended the task to video tracklets in 2016.11 Later datasets added occlusion, infrared imagery, larger galleries, domain shift, and people who were absent from the gallery altogether.

-

The dataset has not disappeared. A CVPR 2026 study of lifelong re-identification still includes VIPeR among seven unseen test datasets, alongside five training datasets.27 Its headline unseen-domain results are averages across those datasets, not individual VIPeR scores. That is another way the question has changed: from learning a matcher for one camera pair to retaining useful representations as new domains arrive.

+

VIPeR hasn't disappeared. A CVPR 2026 study of lifelong re-identification still includes it among seven unseen test datasets, alongside five training datasets.27 Its headline unseen-domain results are averages across those datasets, not individual VIPeR scores. That's another way the question has changed: from learning a matcher for one camera pair to retaining useful representations as new domains arrive.

VIPeR was superseded, which is a good outcome for an early benchmark. It helped establish a measurable problem and then revealed what the next evaluations needed to include. Its small, carefully cropped images remain useful for studying a particular matching task, but high performance there cannot establish how a system behaves in a city, with detection errors, changing clothes, and an unknown population.

-

A promising career, and a question I could not set aside

+

Those small crops can still be hard to match. In 2011, Cheng and colleagues asked 18 people to find matches among 45 candidates. They got their first choice right about 75 percent of the time.13 The pairs had been picked for difficulty, and there were fewer candidates than in the typical 316-image gallery. It shows that people can struggle with these crops too; it doesn't tell us what they'd score on the larger test. When we annotated VIPeR, we could watch frames before and after a person appeared. The released crops contain less information: knowing the correspondence from the footage does not mean the two still images make it unambiguous.

+ +

While writing this retrospective, I tried Claude on VIPeR with help from an AI coding agent. After a little prompt tuning, Haiku 5.5 put the correct match first about 50 percent of the time. Asking Opus 5.5 to rerank its top 30 candidates brought that to about 60 percent, without fine-tuning either model.14 That's below the strongest specialist results, but impressive for so little task-specific work. Given what today's frontier models can do, I'd expected Claude to have an easier time with these images. Twenty years later, VIPeR still gave it plenty of trouble.

+ +

Even so, I hadn't needed to train a dedicated re-identification model; I could ask a general-purpose assistant to do the matching. Deciding when that capability should be used isn't just a choice for researchers who set out to build surveillance systems.

+ +

A promising career, and a question I couldn't set aside

-

I started at UC Santa Cruz about a week after September 11, 2001. By the time I began learning computer vision in 2004, security was widely treated as a reason to accept less privacy. Surveillance research attracted funding and energy, and tracking people across cameras seemed like an important technical problem to help solve.

+

I started at UC Santa Cruz about a week after the terrorist attacks of September 11, 2001. People were scared, and preventing another attack felt urgent. When I began learning computer vision in 2004, surveillance research was often pitched as a way to protect airports and critical infrastructure. In that atmosphere, the argument for trading some privacy for security was persuasive. Tracking people across cameras seemed like a technical problem with an important public purpose.

-

By the time VIPeR was gaining attention, I had several reasons to stay in the area. I understood the problem, had built a benchmark the community wanted, and could see many avenues for further research. The technology had growing commercial demand as well as academic interest. Doubling down would have been a straightforward way to build on the work I had already done.

+

By the time VIPeR was gaining attention, staying in the field would have made sense. I knew the problem, I'd built a benchmark people wanted to use, and there was plenty left to work on. Companies wanted this technology too. I could have kept building on what I'd already done.

But a few years into the Ph.D., I began questioning the application. Before advancing to candidacy in 2008, I changed my proposed dissertation title to Privacy Preserving Pedestrian Recognition. I was trying to find a way to continue the research while addressing the concern that it would make people easier to track.

-

The concern became more concrete through my adviser, Hai Tao, who was building Vion, a company selling surveillance software. My understanding at the time was that its biggest customer was the Chinese government. I could see a path from our work on recognizing people across cameras to systems for tracking people across cities. I did not want that to be the purpose of my research.

+

The concern became more concrete because Hai was also building Vion, a surveillance software company. My understanding at the time was that its largest customer was the Chinese government. This wasn't an unusual direction for computer vision then; much of the field's funding and commercial momentum pointed toward surveillance. But seeing that path up close, from matching people across two cameras to tracking them across a city, made the question hard for me to treat as abstract. I didn't want that to be the purpose of my own work.

-

Vion's later public materials give some context for that concern. The company describes applications in electronic policing, public safety, transportation, and smart-city management. A 2024 article credits Hai and me with VIPeR and describes re-identification as reconstructing movement trajectories. It also presents avoiding face, iris, and gait recognition as a privacy benefit.1314 Those materials do not establish where this particular dataset was used. They do illustrate the distinction I was struggling with: avoiding a particular biometric does not answer whether someone should be tracked.

+

I could propose technical safeguards, but the decision to preserve privacy depended on policy. Avoiding a particular biometric didn't answer whether someone should be tracked. Who could search the camera network? Whose movements could be reconstructed? For what purpose, with what oversight, and with what possibility of refusal? Better features or a formal privacy guarantee could answer only part of that. This became one of several reasons I never finished the Ph.D., and probably one of the larger ones.

-

I could propose technical safeguards, but the decision to preserve privacy depended on policy. Who could search the camera network? Whose movements could be reconstructed? For what purpose, with what oversight, and with what possibility of refusal? Better features or a formal privacy guarantee could answer only part of that. This became one of several reasons I never finished the Ph.D., and probably one of the larger ones.

+

I went on to spend much of my career working on visual search at Amazon. Many of the technical questions were familiar: recognizing things across viewpoints, learning representations, and finding a match in a large gallery. But I was matching products rather than people. The research interests remained; the purpose of the work was different.

-

The current debate about Flock cameras makes that question feel familiar. Reading a license plate is, at its core, optical character recognition: recovering an explicit identifier that is already printed on a vehicle. It has its own difficult imaging conditions, and Flock also searches vehicle characteristics beyond the plate.15 But I do not walk around the city with a license plate on. Person re-identification tries to make someone searchable across cameras using appearance alone. In that sense it could provide a license-plate-like tracking capability for people who carry no such identifier, with the potential to be even more intrusive.

+

The current debate about Flock cameras brings me back to the policy question: who gets to decide whether a community should be tracked? Reading a license plate is, at its core, optical character recognition: recovering an explicit identifier that is already printed on a vehicle. It has its own difficult imaging conditions, and Flock also searches vehicle characteristics beyond the plate.15 But I don't walk around the city with a license plate on. Person re-identification tries to make someone searchable across cameras using appearance alone. In that sense it could provide a license-plate-like tracking capability for people who carry no such identifier, with the potential to be even more intrusive.

-

Last month, in September 2026, my own neighborhood, Ladera in Portola Valley, voted to end its Flock camera contract because of privacy concerns. That brought the question out of research papers and into the place where I live. The decision was about whether we wanted this form of surveillance in our community, not whether the cameras could read plates more accurately.

+

In September 2026, my own neighborhood, Ladera in Portola Valley, voted to end its Flock camera contract because of privacy concerns. That brought the question out of research papers and into the place where I live. The decision was about whether we wanted this form of surveillance in our community, not whether the cameras could read plates more accurately.

-

Looking back, I am proud that VIPeR helped people measure a difficult problem instead of avoiding it. That does not resolve my doubts about what we made easier to do. The question that got me started was whether we could recognize people across cameras. The question I could not set aside was what would happen when we could. The vote in Ladera reminded me that this is not something researchers get to decide for the people who live with the cameras.

+

Looking back, I'm proud that VIPeR helped people measure a difficult problem instead of avoiding it. That doesn't resolve my doubts about what we made easier to do. The question that got me started was whether we could recognize people across cameras. The question I couldn't set aside was what would happen when we could. As researchers, we decide what we make easier. The people who live with those systems should get to decide whether they want them.

VIPeR dataset download, documentation, and citation guidelines

Sources and notes

+

These references cover the text and the methods labeled in the performance chart.

  1. N. Gheissari et al., Person Reidentification Using Spatiotemporal Appearance, CVPR 2006.
  2. D. Gray, S. Brennan, and H. Tao, Evaluating Appearance Models for Recognition, Reacquisition, and Tracking, PETS 2007.
  3. @@ -198,30 +151,31 @@

    Sources and notes

  4. M. Farenzena et al., Person Re-Identification by Symmetry-Driven Accumulation of Local Features, CVPR 2010.
  5. M. Köstinger et al., Large Scale Metric Learning from Equivalence Constraints, CVPR 2012, Table 2(a); authors' code and documentation.
  6. S. Liao et al., Person Re-Identification by Local Maximal Occurrence Representation and Metric Learning, CVPR 2015; authors' results.
  7. -
  8. T. Matsukawa et al., Hierarchical Gaussian Descriptor for Person Re-Identification, CVPR 2016.
  9. +
  10. T. Matsukawa et al., Hierarchical Gaussian Descriptor for Person Re-Identification, CVPR 2016. Labeled GOG + XQDA in the chart.
  11. W. Li et al., DeepReID: Deep Filter Pairing Neural Network for Person Re-identification, CVPR 2014.
  12. L. Zheng, Y. Yang, and A. G. Hauptmann, Person Re-Identification: Past, Present and Future, 2016. See also Karanam et al.'s systematic benchmark and Ye et al.'s deep-learning survey.
  13. L. Zheng et al., Scalable Person Re-Identification: A Benchmark, ICCV 2015.
  14. L. Zheng et al., MARS: A Video Benchmark for Large-Scale Person Re-Identification, ECCV 2016.
  15. Google Scholar article histories for PETS 2007 and ECCV 2008, observed October 12, 2026. The plotted range is 2010–2026, the common range displayed in the two histories.
  16. -
  17. Vion Technology, article on its re-identification patents and history, 2024 (Chinese).
  18. -
  19. Vion Technology, company history; Beijing Security and Protection Industry Association, company profile (Chinese).
  20. +
  21. D. S. Cheng et al., Custom Pictorial Structures for Re-identification, BMVC 2011, Section 4.3 and Figure 6(e). Eighteen people completed 180 trials on 45 selected pairs; the approximate first-choice accuracy is read from the plot. The legend's 98.56 is normalized CMC area, not rank-one accuracy. The often-cited 71.08 percent in Liu et al.'s POP study, ICCV 2013, Table 2, is a human-assisted ranking result after three feedback rounds, not unaided human accuracy.
  22. +
  23. Exploratory experiment completed in October 2026: one frozen 316-identity test half, both camera directions, no VIPeR weight training. After excluding identities used for prompt development, Haiku scored 296 of 606 queries correctly at rank one (48.8 percent); top-30 Opus reranking scored 355 (58.6 percent). Three malformed rankings count as failures. This is one split, not a five-repeat estimate or a blind test; earlier diagnostics inspected some evaluation images, and foundation-model pretraining exposure is unknown. Method, prompts, and measured results.
  24. Flock Safety, Vehicle Recognition Software for Investigation-grade LPR, 2026; describes plate reading and searches by vehicle appearance.
  25. R. Zhao, W. Ouyang, and X. Wang, Unsupervised Salience Learning for Person Re-identification, CVPR 2013, Table 1. The chart uses eSDC ocsvm with feature fusion.
  26. -
  27. R. Zhao, W. Ouyang, and X. Wang, Person Re-identification by Salience Matching, ICCV 2013, Section 5 and Figure 5(a).
  28. -
  29. Y. Yang et al., Salient Color Names for Person Re-identification, ECCV 2014, Table 1. The chart uses Final(ImgF), including color-histogram fusion.
  30. -
  31. S. Paisitkriangkrai, C. Shen, and A. van den Hengel, Learning to Rank in Person Re-identification with Metric Ensembles, CVPR 2015, Figure 1.
  32. -
  33. J. García et al., Person Re-Identification Ranking Optimisation by Discriminant Context Information Analysis, ICCV 2015, Table 2. Includes gallery-context re-ranking.
  34. -
  35. L. Zhang, T. Xiang, and S. Gong, Learning a Discriminative Null Space for Person Re-Identification, CVPR 2016, Table 1, Ours (Fusion).
  36. -
  37. X. Jin et al., Style Normalization and Restitution for Generalizable Person Re-Identification, CVPR 2020, supplementary Table 5. The five-source setting includes DukeMTMC-reID; the alternative four-source setting reports 55.1 percent.
  38. -
  39. Y. Dai et al., Generalizable Person Re-Identification with Relevance-Aware Mixture of Experts, CVPR 2021, Table 2, Protocol 1. Protocol 2 uses different source datasets and reports 63.4 percent.
  40. -
  41. P. Zhang et al., Adaptive Cross-Domain Learning for Generalizable Person Re-Identification, ECCV 2022, Table 1. Uses four source datasets, omitting DukeMTMC.
  42. -
  43. R. Nie et al., Rethinking Normalization Layers for Domain Generalizable Person Re-Identification, ECCV 2024, Table 1.
  44. +
  45. R. Zhao, W. Ouyang, and X. Wang, Person Re-identification by Salience Matching, ICCV 2013, Section 5 and Figure 5(a). Labeled SalMatch in the chart.
  46. +
  47. Y. Yang et al., Salient Color Names for Person Re-identification, ECCV 2014, Table 1. The chart labels this SCNCD and uses Final(ImgF), including color-histogram fusion.
  48. +
  49. S. Paisitkriangkrai, C. Shen, and A. van den Hengel, Learning to Rank in Person Re-identification with Metric Ensembles, CVPR 2015, Figure 1. Labeled Metric ensembles in the chart.
  50. +
  51. J. García et al., Person Re-Identification Ranking Optimisation by Discriminant Context Information Analysis, ICCV 2015, Table 2. KCCA + DCIA includes gallery-context re-ranking.
  52. +
  53. L. Zhang, T. Xiang, and S. Gong, Learning a Discriminative Null Space for Person Re-Identification, CVPR 2016, Table 1, Ours (Fusion), labeled NFST fusion in the chart.
  54. +
  55. X. Jin et al., Style Normalization and Restitution for Generalizable Person Re-Identification, CVPR 2020, supplementary Table 5. The chart uses SNR's five-source setting, including DukeMTMC-reID; the alternative four-source setting reports 55.1 percent.
  56. +
  57. Y. Dai et al., Generalizable Person Re-Identification with Relevance-Aware Mixture of Experts, CVPR 2021, Table 2, Protocol 1, labeled RaMoE in the chart. Protocol 2 uses different source datasets and reports 63.4 percent.
  58. +
  59. P. Zhang et al., Adaptive Cross-Domain Learning for Generalizable Person Re-Identification, ECCV 2022, Table 1. ACL uses four source datasets, omitting DukeMTMC.
  60. +
  61. R. Nie et al., Rethinking Normalization Layers for Domain Generalizable Person Re-Identification, ECCV 2024, Table 1. Labeled ReNorm in the chart.
  62. H. Zhao, L. Qi, and X. Geng, CILP-FGDI: Exploiting Vision-Language Model for Generalizable Person Re-Identification, IEEE TIFS 2025, Table I; authors' implementation. The title spells the acronym CILP-FGDI; the method and code use CLIP-FGDI. The chart uses the main-table 76.0 percent, not the 77.5 percent hyperparameter-ablation result in Table VIII.
  63. K. Xu et al., Vision-Language Attribute Disentanglement and Reinforcement for Lifelong Person Re-Identification, CVPR 2026; supplementary Tables 3-4. UnSeen-Avg combines seven datasets and is not a VIPeR rank-one result.
  64. J. Song et al., Generalizable Person Re-Identification by Domain-Invariant Mapping Network, CVPR 2019, Table 2.
  65. J. Jia, Q. Ruan, and T. M. Hospedales, Frustratingly Easy Person Re-Identification: Generalizing Person Re-ID in Practice, BMVC 2019, Table 5. The chart uses ResNet50 + DualNorm; the main MobileNetV2 configuration reports 53.9 percent.
  66. -
  67. K. Zhou et al., Learning Generalisable Omni-Scale Representations for Person Re-Identification, October 2019 preprint, Table 4. Includes VIPeR fine-tuning after external re-identification pretraining.
  68. +
  69. K. Zhou et al., Learning Generalisable Omni-Scale Representations for Person Re-Identification, October 2019 preprint, Table 4. The chart's OSNet result includes VIPeR fine-tuning after external re-identification pretraining.
  70. +
  71. Q. Zhao and H. Tao, Object Tracking Using Color Correlogram, IEEE VS-PETS 2005, pp. 263-270.