Occluded and Partial Person ReID: Methods and Limits

At the edge of a crowded dance-floor photograph, you can see a sequined sleeve, half a torso and one hand. A partner covers the shoulder; the frame cuts off everything below the waist. The gallery image, meanwhile, shows a full body from behind. What should the model compare?

That is the tension behind occluded and partial Person ReID. Occlusion hides appearance behind another person or object. A partial query contains only part of the body. Both break the convenient assumption that query and gallery offer comparable, aligned evidence.

A global model may compress the visible torso, the occluder and the background into one vector. Occlusion-aware research asks narrower questions:

• Which regions are actually visible?
• Which visible query region corresponds to which gallery region?
• Should a covered region contribute less to the final similarity?
• Can training prevent the model from relying on one easy cue?

The answer is never “infer the hidden person with certainty.” It is “make the comparison less dependent on unavailable or misleading pixels.” No architecture can recover a shoe, jacket or body region that is absent from every image.

For the basic retrieval setup and its limits, see the Person ReID guide.

Occluded and partial are related, not identical

Occlusion means something blocks the person inside the frame. The crop may still have the expected size, but a chair, another participant or foreground object covers some regions.

Truncation means the image boundary or detector box cuts off a region. There are no pixels for the missing lower body or side.

Partial ReID is a retrieval setting in which the query intentionally or unavoidably contains only a portion of the person. The gallery may contain holistic images.

Misalignment occurs when nominally corresponding spatial regions contain different anatomy because pose, crop position or scale changed.

These states can overlap. A partial crop can also be occluded and misaligned. An evaluation should label them separately because a method that handles fixed truncation may still fail when an occluder contributes a strong misleading texture.

Keep local evidence from disappearing into the average

PCB divides a spatial feature map into uniform horizontal stripes and learns a descriptor for each stripe. This prevents every local cue from being averaged immediately into one global vector.

The idea is simple and useful, but a stripe is not an anatomical detector. A top stripe might contain a head in one crop and mostly background in another. A bent pose can move the same costume region vertically. Fixed stripes are therefore a baseline for locality, not a visibility solution by themselves.

Multi-granularity designs add coarser and finer partitions so that a missing small region does not necessarily invalidate the entire descriptor. They also create more opportunities for an occluder to contaminate one branch, which is why matching rules matter.

Compare what both images can actually see

The Visibility-Aware Part Model predicts both part features and whether those parts are visible. Distance can then be computed over parts visible in both images rather than treating missing parts as negative evidence.

This establishes an important pattern: feature extraction and visibility estimation are distinct tasks. A reliable local descriptor is not enough if the system does not know whether that descriptor came from the person, an occluder or padding.

Visibility predictions can also be wrong. They depend on the training supervision and domain. Clothing, unusual poses and tight event crops may differ from benchmark imagery.

Use pose to put shoulders beside shoulders

Pose-Guided Feature Alignment uses pose landmarks to identify visible body regions and align part features. Pose offers semantic guidance: a visible shoulder can be compared with a shoulder instead of whatever occupies the same horizontal coordinate.

Pose-Guided Visible Part Matching similarly focuses matching on shared visible parts. These methods show that correspondence matters, not just local pooling.

Their dependency is also clear. Pose estimation can fail on the very images that make occluded ReID difficult: motion blur, extreme articulation, overlapping people and severe truncation. The cost and error profile of the pose model belong in the end-to-end evaluation.

Hide easy cues during training

Batch DropBlock drops the same contiguous feature-map region across all images in a batch. This encourages the model to discover additional discriminative regions rather than repeatedly relying on its easiest cue.

Feature dropping is related to occlusion robustness, but simulated missing features are not a complete model of real occlusion. A real occluder contributes its own pixels and may resemble a person. Dropping a clean rectangle does not reproduce complex boundaries, detector errors or two overlapping identities.

Let the model weight its regional evidence

An internal DINO-ReID report describes a confidence-weighted part head. It divides a backbone feature map into horizontal stripes, pools each stripe, estimates stripe confidence from mean absolute activation, normalizes those scores within the image, and applies a temperature-controlled softmax before combining the part descriptors. It also uses stripe dropout during training.

The report presents that activation-based weighting and training combination as an internal contribution. Horizontal stripes come from established part-based ReID; spatial feature dropping and visibility-aware matching also predate it. A complete prior-art review would be needed to confirm novelty even for the precise combination. It should not be described as the invention of occlusion-aware or part-based ReID.

Activation magnitude is not ground-truth visibility. A bright background or occluder may create a high activation. A genuinely useful but subtle garment region may have a lower score. Min-max normalization can also be unstable when all stripe scores are similar. These are hypotheses to test with controlled occlusion labels and ablations.

Five strategies, five dependencies

Method family	Extra signal	How it handles missing evidence	Main dependency
Fixed part pooling	Spatial position	Keeps regional descriptors separate	Reasonably aligned crops
Visibility-aware parts	Predicted visibility	Excludes or downweights unavailable parts	Accurate visibility prediction
Pose-guided alignment	Keypoints or pose maps	Matches semantically corresponding visible regions	Robust pose estimation
Feature dropping	Training-time masks	Discourages reliance on one region	Realistic drop policy
Activation confidence	Internal feature magnitude	Weights stripes dynamically	Confidence must correlate with useful visibility

No row is universally strongest. A hybrid can combine several rows, which makes one-factor ablation even more important.

Build a test that makes occlusion visible

• Define whether each example is occluded, truncated, partial, misaligned or multi-person.
• Keep identity splits disjoint between training, calibration and test.
• Record visible-area ranges and who created those annotations.
• Include holistic-to-holistic, partial-to-holistic and partial-to-partial retrieval.
• Slice results by occluder type, pose, crop boundary and blur.
• Compare against the same holistic baseline with identical training data.
• Remove one visibility, pose or confidence component at a time.
• Report raw embedding retrieval separately from re-ranking.
• Inspect false matches where the occluder dominates the crop.
• Measure abstention and reviewer workload, not only Rank-1 and mAP.

Before embedding, the input pipeline should also reject crops that contain the wrong primary subject. The person crop quality checklist covers that separate engineering boundary.

What the remaining pixels cannot prove

The cited papers do not prove that a general ReID model is automatically robust to occlusion. They evaluate particular methods under particular benchmark protocols. A strong result on one synthetic or curated occlusion setting does not guarantee performance with overlapping dancers, unusual costumes or a new detector.

They also do not prove that horizontal activation is a true visibility measurement. That is an internal design hypothesis. Without a controlled baseline that changes only the weighting mechanism, a gain cannot be assigned to confidence weighting rather than the backbone, training data, losses or augmentation.

A private occlusion experiment also says nothing about a served revision unless the checkpoint, preprocessing and test protocol match. Even a validated occlusion-aware embedding returns similarity candidates; it does not establish civil identity or recreate evidence that was never visible.

Sources

• Beyond Part Models: Person Retrieval with Refined Part Pooling
• Perceive Where to Focus: Learning Visibility-Aware Part-Level Features for Partial Person Re-Identification
• Pose-Guided Feature Alignment for Occluded Person Re-Identification
• Batch DropBlock Network for Person Re-Identification and Beyond
• Pose-Guided Visible Part Matching for Occluded Person ReID