Source-linked AI summary
Batch DropBlock Network for Person Re-identification and Beyond
Zuozhuo Dai, Mingqiang Chen, Xiaodong Gu, Siyu Zhu, Ping Tan
TL;DR
Person re-identification is challenged by pose changes and occlusions that can suppress attentive local features. BDB combines global and feature-dropping branches, synchronously erases feature regions during training, and concatenates both representations. Experiments report improvements on person re-identification and image-retrieval benchmarks, including 72.8% Rank-1 accuracy on CUHK03-Detect, 6.0% above the state-of-the-art work.
Problem
Pose, illumination, and viewpoint changes can cause CNNs to suppress discriminative local body parts in person re-identification.
Method
BDB combines a global branch with a feature-dropping branch using Batch DropBlock, which drops the same feature region across a training batch before concatenating branch features.
Results
72.8% Rank-1 accuracy on CUHK03-Detect was 6.0% higher than the state-of-the-art work, while BDB consistently improved various metric-learning schemes on image-retrieval datasets.
Takeaways & Limitations
The BDB Network provides a generalized feature-learning mechanism applicable to person re-identification and other metric-learning tasks.
Takeaways & Limitations
Experiments use a single-query setting without re-ranking.
Abstract
from arXiv · showhide
Since the person re-identification task often suffers from the problem of pose changes and occlusions, some attentive local features are often suppressed when training CNNs. In this paper, we propose the Batch DropBlock (BDB) Network which is a two branch network composed of a conventional ResNet-50 as the global branch and a feature dropping branch. The global branch encodes the global salient representations. Meanwhile, the feature dropping branch consists of an attentive feature learning module called Batch DropBlock, which randomly drops the same region of all input feature maps in a batch to reinforce the attentive feature learning of local regions. The network then concatenates features from both branches and provides a more comprehensive and spatially distributed feature representation. Albeit simple, our method achieves state-of-the-art on person re-identification and it is also applicable to general metric learning tasks. For instance, we achieve 76.4% Rank-1 accuracy on the CUHK03-Detect dataset and 83.0% Recall-1 score on the Stanford Online Products dataset, outperforming the existing works by a large margin (more than 6%).
1. Introduction
The paper addresses suppressed local features in person re-identification by introducing BDB, which combines global representations with attentive local feature learning through synchronized feature dropping.
- Pose, illumination, and viewpoint changes challenge CNN-based person re-identification, while unstable body parts can be suppressed in learned representations.
- BDB is a two-branch network with a conventional global branch and a feature dropping branch for roughly aligned metric learning tasks.
- Concatenating both branches produces more comprehensive, spatially distributed representations instead of relying on only a few discriminative features.
- Batch DropBlock drops the same feature-map region across all images in a training batch, reinforcing attention to the remaining local parts.
- 72.8% Rank-1 accuracy on CUHK03-Detect was 6.0% higher than the state-of-the-art work.
2. Related work
Related work frames person re-identification as metric learning under substantial visual variation, with prior methods using losses, body-part structure, attention, and augmentation.
- Person re-identification must handle large variation in pose, background, illumination, and camera conditions.
- Metric-learning research develops embedding features and specialized objectives such as contrastive, triplet, lifted-structure, quadruplet, and histogram losses.
- Part-based methods aggregate body-part and global cues, often by horizontally splitting feature maps into fixed strips.
- Attention and saliency methods address imperfect detection and body-part misalignment by focusing on attentive regions.
- The proposed architecture is illustrated for person re-identification with global and feature-dropping branches.
3. Batch DropBlock (BDB) Network
The BDB Network combines a ResNet-50 global branch with a feature-dropping branch that synchronously erases feature regions, then concatenates both embeddings for testing.
- Network Architecture: The network uses ResNet-50 as its backbone and appends a global branch after Stage 4.
- Batch DropBlock Layer: The Batch DropBlock Layer randomly zeroes the same large region of tensor T for every feature map in a batch.
- Network Architecture: The global branch supervises the feature-dropping branch and helps it learn from a well-learned feature map.
- Network Architecture: The feature dropping branch applies Batch DropBlock, uses global max pooling, and reduces its feature vector from 2048 to 1024 dimensions.
- Pooling: Global average pooling is used in the global branch, while global max pooling encourages weaker salient features after dominant features are dropped.
- Network Architecture: During testing, features from both branches are concatenated as the pedestrian-image embedding, while the Batch DropBlock layer adds no parameters or network size.
- Training: Training uses soft-margin batch-hard triplet loss and softmax loss on both branches.
4. Experiments
Experiments evaluate BDB on person re-identification and standard image-retrieval datasets using multiple metric-learning loss functions.
- BDB is evaluated on benchmark person re-identification datasets and standard image-retrieval datasets with different metric-learning losses.
4.1. Person re-ID Experiments
The BDB Network is evaluated on person re-identification benchmarks and through ablations of its branches, dropout design, pooling, loss, and erasing ratio. It achieves strong benchmark performance while both branches and Batch DropBlock contribute to results.
- Comparison with State-of-the-Art: The BDB Network achieves state-of-the-art performance on CUHK03 and DukeMTMC-reID, while matching MGN on Market-1501 with one-third its parameter count.The largest improvement over prior methods occurs on CUHK03-Detect, described as the most challenging dataset.
- Comparison with State-of-the-Art: Given a back-view query, BDB Network retrieves front-view and side-view images of the same person.
- Ablation Studies: Adding the global branch improves performance beyond the feature-dropping branch alone, indicating that both branches contribute to the final representation.The branches capture salient appearance clues and fine-grained discriminative features.
- Ablation Studies: Batch DropBlock is more effective than the evaluated Dropout and DropBlock alternatives on the CUHK03-Detect dataset.Unlike the alternatives, it drops a large contiguous region shared across feature maps within a batch.
- Ablation Studies: Triplet loss significantly improves BDB Network performance, because Batch DropBlock operates by considering relationships between images.
- Ablation Studies: The best performance is achieved with an erased height ratio of 0.3 while the erased width ratio is fixed at 1.0.The same 0.3 height-ratio setting is used in the person re-identification experiments.
4.2. Image Retrieval Experiments
The BDB Network is evaluated on roughly aligned image-retrieval datasets and achieves strong Recall@1 performance across both small- and large-scale benchmarks. Its two-branch representation also provides attentive detail features and generalizes across standard metric-learning losses.
- 4.2.1 Datasets and Settings: The evaluation uses CUB200-2011, CARS196, Stanford Online Products, and In-Shop Clothes Retrieval datasets, with cropped inputs for the roughly aligned CUB200-2011 and CARS196 settings.The BDB Network requires input images to be roughly aligned.
- 4.2.2 Comparison with State-of-the-Art: The BDB Network achieves the best Recall@1 scores on all experimental image-retrieval datasets.The comparison is reported in Table 7 against other state-of-the-art metric-learning methods.
- 4.2.2 Comparison with State-of-the-Art: +3.5% improvement is achieved on the small-scale CUB200-2011 dataset, described as the most challenging dataset in the comparison.
- 4.2.2 Comparison with State-of-the-Art: 6.7% is the BDB Network’s improvement over the state-of-the-art on Stanford Online Products, which contains 22,634 classes and 120,053 product images.
- 4.2.2 Comparison with State-of-the-Art: The two-branch network encodes comprehensive features with attentive detail features, helping explain robustness to illumination, pose, and occlusion variance.
5. Conclusion
The paper proposes Batch DropBlock and the BDB Network to improve neural-network training for person re-identification and general metric-learning tasks. Experiments on person-re-identification and image-retrieval datasets report significant improvements across these benchmarks.
- 5. Conclusion: Batch DropBlock is proposed to improve neural-network optimization for person re-identification and general metric-learning tasks.
- 5. Conclusion: The BDB Network combines a global branch for salient representations with a feature-erasing branch for detailed features.
- 5. Conclusion: Extensive experiments on person-re-identification and image-retrieval datasets show significant improvements on person-re-identification and general image-retrieval benchmarks.