Academic guide8 min read

Person ReID Model Architectures: CNNs, Transformers, Parts

Compare CNN, transformer, part-based and hybrid Person ReID model architectures, including pooling choices, trade-offs and evaluation criteria.

PolyReID Engineering25 July 2026

Send the same dancer crop through two ReID systems. One moves through stacked convolutional filters; the other lets image patches exchange information through attention. Both eventually hand you a row of numbers. From the API response alone, the architectural argument has disappeared—but its effects on speed, memory and failure cases have not.

A Person ReID model architecture is the path from that crop to a compact appearance vector. Three design families appear repeatedly in the literature: convolutional neural networks (CNNs), vision transformers, and heads that preserve body regions or multiple granularities.

Which one should you choose? Not the one with the most impressive name. A small CNN may fit a tight latency budget. A transformer can model relationships between distant patches, but may cost more to run. A part-based head can preserve a visible sleeve or shoe, yet fixed horizontal parts suffer when the crop is tilted or badly aligned. “Best” only becomes meaningful after you name the data, hardware and errors that matter.

If the retrieval task itself is unfamiliar, start with the Person ReID guide. It separates person detection, embedding generation, nearest-neighbour search and human review.

Three ways to read the same crop

Each family summarizes the same crop differently:

A CNN scans for patterns such as edges, colors and textures, gradually combining nearby evidence into larger visual concepts.
A vision transformer splits the crop into patches and lets those patches exchange information through attention.
A part-based model keeps separate summaries for regions such as upper and lower portions of the feature map instead of compressing everything immediately.

These are not mutually exclusive camps. You can use a transformer backbone, pool one global vector, create several part vectors and fuse them into one embedding. “CNN,” “transformer” and “part-based” describe choices inside a system, not sealed product categories.

Follow one crop through four blocks

Whatever family you choose, most embedding models can be read as four blocks:

Input preprocessing resizes and normalizes a one-person crop.
A backbone creates a spatial feature map or a sequence of patch tokens.
A pooling or aggregation head converts variable spatial evidence into one or more fixed-size vectors.
A training head and losses shape the feature space so same-identity training samples tend to be close and different identities tend to be separated.

The identity classifier used during training may be discarded at inference. The remaining backbone and embedding head do not output a person’s name. They output coordinates in a learned appearance space.

Pooling is an architectural decision too. Average pooling treats all locations equally, max pooling keeps only the strongest activation per channel, and generalized mean pooling, or GeM, learns a position between those behaviors. The original GeM image-retrieval paper introduced a trainable pooling exponent. Reusing GeM in Person ReID is an application of that published operator, not a new pooling invention.

CNNs: build a representation from nearby patterns

CNNs encode a strong locality assumption: nearby pixels are processed together, and repeated filters are applied across the image. This makes them efficient and well suited to visual patterns that recur at different positions.

OSNet is a useful ReID-specific example. Its omni-scale block processes features through streams with different effective receptive fields, then learns how to combine them. The aim is to capture both small details and larger structures in a lightweight network. That is more specific than saying “OSNet is a small CNN”: its contribution is the multi-scale feature design and adaptive aggregation.

Common CNN advantages include mature tooling, predictable memory use and good throughput on many devices. Their limitations are not absolute, but long-range relationships must emerge through stacked local operations. Downsampling can also remove fine details, while a global pooling head can hide which region supplied the useful evidence.

Transformers: let distant patches exchange context

A vision transformer represents an image as patch tokens. Self-attention lets every token combine information from other tokens, so distant regions can interact directly. Positional information is required because attention alone does not know whether a patch came from the shoulder or the shoes.

TransReID established a transformer-based object ReID baseline and proposed two additions: side-information embeddings for factors such as camera or viewpoint, and a jigsaw patch module that creates local feature groups. A model that merely uses a transformer backbone does not automatically implement either addition.

Self-supervised backbones introduce another distinction. DINOv2, released in 2023, learns general-purpose visual features without ReID identity labels. DINOv3 is a separate Meta technical report from August 2025 with a different large-scale training recipe, including Gram anchoring to preserve dense feature quality during long training. A citation to DINOv2 cannot substantiate a claim specifically about DINOv3. Neither report, by itself, proves performance on Person ReID, ballroom photography or any private event dataset.

Transformers can retain rich patch-level evidence, but attention is not free. Image resolution, patch size, token count, training data and inference hardware all affect cost. A generic foundation backbone also still needs an evaluated ReID head and training objective.

Part-based methods: do not average every clue away

Global pooling is convenient, but one vector can be dominated by a bright jacket, a large background region or another strong cue. Part-based models try to preserve multiple pieces of evidence.

PCB divides a convolutional feature map into uniform horizontal stripes and learns a descriptor for each stripe. Its refined part pooling component was designed to improve the consistency of those assigned regions. The stripes are not detected anatomical body parts; they are spatial bands whose meaning depends on reasonably aligned person crops.

MGN, the Multiple Granularity Network, combines a global branch with local branches divided at different granularities. It therefore collects whole-person context and increasingly local evidence. Describing another model as “inspired by global and local branches” is not the same as saying it implements MGN’s exact branch structure, losses or inference descriptor.

Part methods can help when some local evidence remains informative, but they do not reconstruct invisible pixels. Severe truncation can remove entire bands. Pose changes can move the same garment across bands. A neighbouring person can occupy a band and become a misleading feature.

Put the trade-offs on one page

Family	Main representation	Typical strength	Typical risk	What to measure
CNN	Hierarchical spatial feature maps	Efficient local and multi-scale pattern extraction	Fine evidence may be lost through downsampling	Latency, memory, exact-search retrieval quality
Transformer	Contextualized patch tokens	Direct long-range patch interaction	Token cost and dependence on training/pretraining choices	Throughput by resolution, domain-sliced ranking
Part-based head	Several regional descriptors	Keeps local evidence separate	Alignment and occlusion can corrupt fixed regions	Performance by truncation, pose and crop quality
Hybrid	Global plus local features over either backbone	Can combine context and detail	More components and tuning choices	Controlled component ablations and end-to-end cost

Do not fill this table with headline scores copied from unrelated papers. Dataset splits, input sizes, pretraining, losses and re-ranking rules can all differ. If you cannot keep those conditions comparable, you are comparing complete experiments—not isolated architectures.

How the internal report maps to prior work

An internal DINO-ReID research report describes a hybrid: a DINOv3 vision-transformer backbone, global GeM pooling, horizontal part pooling, a learned global/part fusion, and loss functions drawn from prior research. Those ingredients should be attributed to their primary sources.

The report’s more specific internal mechanism estimates a confidence for each stripe from activation magnitude, normalizes the stripe scores, applies a temperature-controlled softmax, and reduces reliance on random stripes during training through stripe dropout. The report presents that precise combination as an internal contribution. That is an authorship claim, not confirmation of literature-wide novelty; the broader concepts of part pooling, spatial dropout and visibility-aware matching predate it.

This distinction matters for technical writing. “We combine published components in a new event-oriented design” is supportable. “We invented part-based ReID” is not. The internal evaluation can motivate a controlled public study, but it does not replace one.

Before you choose, write down the test

Before selecting a family, record:

the intended crop quality, resolution and aspect-ratio policy;
the deployment latency, throughput and memory budget;
how often people are occluded, truncated or badly aligned;
whether the test domain is genuinely unseen during training;
the exact pretraining source and permitted data use;
which descriptor is indexed and whether it is normalized;
whether any re-ranking step is included in reported metrics;
results from the same train, validation and test protocol;
component ablations that change one architectural factor at a time;
failure slices and reviewer workload, not only average Rank-1 or mAP.

Architecture and optimization also interact. Change the loss and the apparent winner can change even when the backbone does not. The companion article on Person ReID loss functions follows that part of the experiment.

Evidence boundaries for architecture claims

No cited architecture paper proves that its design is best for every ReID deployment. Published scores are evidence for the paper’s stated datasets, protocol and implementation—not universal rankings between CNNs and transformers.

Likewise, the internal report does not contain a controlled same-data, same-loss, same-compute ablation establishing that DINOv3 is superior to OSNet. Changing the backbone, training identities and other components together cannot isolate the cause of a gain. DINOv3’s self-supervised pretraining is not itself proof of occlusion robustness or cross-event generalization.

Before internal results are used to describe a served system, the evaluated and served checkpoints, preprocessing, descriptor, gallery protocol and test set must be traceable. An architecture name alone cannot establish that equivalence.