Diverse Part Discovery: Occluded Person Re-identification with Part-Aware Transformer

• Occlusion hides parts of pedestrians in real-world scenes, making it essential to learn discriminative features from visible regions while avoiding noisy occluded regions.Existing approaches use hand-crafted splits, external parsing or pose models, or attention; these can introduce background noise, fail under severe occlusion, or focus on only the most discriminative region.Source: §1 · pp. 1–2, Figure 1 · p. 1
• PAT uses a pixel-context transformer encoder and part-prototype transformer decoder to discover human parts from identity labels, forming an end-to-end model for occluded Re-ID.The encoder models pixel correlations to reduce background clutter; learnable prototypes act as decoder queries to generate part-aware masks and pooled part features.Source: §1 · p. 2
• Part diversity encourages different part features to focus on distinct foreground regions, while part discriminability preserves identity information through classification and triplet loss.The mechanisms guide prototype learning without part annotations, enabling weakly supervised discovery across the dataset.Source: §1 · p. 2
• The paper evaluates PAT on occluded, partial, and holistic Re-ID across six standard datasets, reporting favorable performance against state-of-the-art methods.The introduction identifies the work as the first unified deep model to exploit a transformer encoder-decoder architecture for occluded person Re-ID.Source: §1 · p. 2

Partial Person Re-Identification

• AMC and Sliding Window Matching, DSR [10], STNReID [29], and VPM [37] address partial probes, but require manual cropping of the occluded target.The paper notes that manual cropping is inefficient in practice and may introduce human bias.Source: §2 · p. 3

Holistic Person Re-Identification

• Holistic Re-ID work uses parsing [19], uniform partitions [38], or attention [51] [26] to learn part features, but assumes complete bodies; PAT adaptively captures discriminative parts.The paper positions PAT's transformer encoder-decoder as an alternative for occluded person Re-ID.Source: §2 · p. 3

Occluded Person Re-Identification

• Prior occluded Re-ID uses occlusion/ID classification [61], reconstruction [13], pose guidance [5], or graph convolution [40], while PAT discovers diverse parts using identity labels alone.Most such methods rely heavily on off-the-shelf human parsing or pose estimators; PAT uses a transformer encoder-decoder in a weakly supervised manner.Source: §2 · p. 3

3.1. Pixel Context based Transformer Encoder

• ResNet-50 features are flattened into pixel tokens that self-attend to aggregate related pixels, yielding context-aware features more robust to background clutter.The backbone output has spatial dimensions H/16 × W/16; the global pooled feature is trained with identity classification and triplet losses.Source: §3.1 · pp. 3–4

3.2. Part Prototype based Transformer Decoder

• Learnable part prototypes use self-attention to exchange context and cross-attend to pixel features, producing part-aware masks and weighted-pooled part features.The prototypes are trained using identity labels without human part annotations.Source: §3.2 · p. 5
• Part diversity discourages prototypes from focusing on the same foreground region, while part classification and triplet losses preserve identity-discriminative part features.The diversity loss expands discrepancies among part features; triplet learning treats same-identity features as positive pairs and different-identity features as negative pairs.Source: §3.2 · p. 5

3.3. Training and Inference

• PAT is trained with identity labels and represents each test image by concatenating its global feature with its part features.The training objective is defined for occluded person Re-ID; the concatenated representation is used for images of unseen identities.Source: §3.3 · p. 5

4.1. Datastes and Evaluation Metrics

• The evaluation spans two occluded, two partial, and two holistic Re-ID datasets, using CMC curves and mAP to assess model quality.The six datasets are Occluded-Duke, Occluded-REID, Partial-REID, Partial-iLIDS, Market-1501, and DukeMTMC-reID.Source: §4.1 · p. 6

4.2. Implementation Details

• PAT uses ImageNet-pretrained ResNet-50, with K=6 on Market-1501 and K=14 on other datasets, and trains all modules jointly for 120 epochs.Images are resized to 256 × 128; training uses batch size 64 with four images per person and a learning rate initialized to 3.5 × 10−4.Source: §4.2 · p. 6

4.3. Comparison with State-of-the-art Methods

• 64.5%/53.6% Rank-1/mAP on Occluded-Duke and 81.6%/72.1% on Occluded-REID are reported as state-of-the-art results for PAT.PAT exceeds PCB on Occluded-Duke by +21.9% Rank-1 accuracy and +19.9% mAP, and exceeds PVPM and HOReID there by at least +9.4% Rank-1 accuracy and +9.8% mAP.Source: Table 1 · p. 6, §4.3 · pp. 6–7
• 88.0%/92.3% Rank-1/Rank-3 on Partial-REID and 76.5%/88.2% on Partial-iLIDS outperform all previous partial Re-ID models in the cross-domain evaluation.PAT exceeds HOReID by +2.7% Rank-1 accuracy on Partial-REID and +3.9% on Partial-iLIDS.Source: Table 2 · p. 7, §4.3 · p. 7
• 95.4%/88.0% Rank-1/mAP on Market-1501 and 88.8%/78.2% on DukeMTMC-reID give PAT competitive performance against state-of-the-art holistic Re-ID methods.PAT exceeds HOReID by +3.1% mAP on Market-1501 and +2.6% mAP on DukeMTMC-reID.Source: Table 3 · p. 7, §4.3 · p. 7

4.4. Ablation Studies

• On Occluded-Duke, the encoder adds +7.1% mAP; adding the part-prototype decoder adds +12.9% to 51.7% mAP, and decoder self-attention adds +0.9% mAP.The part diversity loss gives the full model its best performance and guides prototypes to discover diverse discriminative human parts; part discriminability is the model's basis.Source: §4.4 · pp. 7–8, Table 4 · p. 8
• On Occluded-Duke, performance improves as K increases up to 14, while Market-1501 performs best at K=6, consistent with the more complex Occluded-Duke scenarios.The authors attribute the difference to the usefulness of finer-grained part features in more complex scenarios.Source: Figure 3 · p. 8, Figure 4 · p. 8, §4.4 · p. 8

4.5. Visualization of Discovered Parts

• Different prototypes capture diverse discriminative parts, and their fused masks span most of the person to reduce background interference and occlusion.In the visualization, the first prototype focuses mainly on the head and the second on the upper body.Source: Figure 5 · p. 8, §4.5 · p. 8

• PAT uses learnable part prototypes to discover diverse, discriminative human parts for occluded person Re-ID.The paper presents PAT as an end-to-end method.Source: §5 · p. 8
• Part diversity and part discriminability mechanisms learn the prototypes using only identity labels in a weakly supervised manner.Source: §5 · p. 8
• Experiments across three tasks on six standard Re-ID datasets demonstrate the effectiveness of PAT.Source: §5 · p. 8