Source-linked AI summary
Combined Depth Space based Architecture Search For Person Re-identification
Hanjun Li, Gaojie Wu, Wei-Shi Zheng
TL;DR
The paper addresses the need for lightweight ReID architectures that are more suitable and computationally efficient than large classification backbones. It searches a Combined Depth Space with Top-k Sample Search, adds FBLNeck to reconcile training losses, and reports state-of-the-art performance among lightweight networks.
Problem
Large classification backbones used for ReID are computationally expensive and may not provide architectures suitable for the task.
Method
The paper searches Combined Depth Space for CDNet using differentiable architecture search and Top-k Sample Search, with FBLNeck balancing triplet and softmax losses.
Results
CDNet achieves state-of-the-art performance on ReID and other tasks among lightweight networks.
Takeaways & Limitations
The proposed design yields a lightweight and efficient network for person re-identification.
Abstract
from arXiv · showhide
Most works on person re-identification (ReID) take advantage of large backbone networks such as ResNet, which are designed for image classification instead of ReID, for feature extraction. However, these backbones may not be computationally efficient or the most suitable architectures for ReID. In this work, we aim to design a lightweight and suitable network for ReID. We propose a novel search space called Combined Depth Space (CDS), based on which we search for an efficient network architecture, which we call CDNet, via a differentiable architecture search algorithm. Through the use of the combined basic building blocks in CDS, CDNet tends to focus on combined pattern information that is typically found in images of pedestrians. We then propose a low-cost search strategy named the Top-k Sample Search strategy to make full use of the search space and avoid trapping in local optimal result. Furthermore, an effective Fine-grained Balance Neck (FBLNeck), which is removable at the inference time, is presented to balance the effects of triplet loss and softmax loss during the training process. Extensive experiments show that our CDNet (~1.8M parameters) has comparable performance with state-of-the-art lightweight networks.
1. Introduction
The paper targets lightweight, ReID-suitable architectures because conventional large classification backbones are computationally costly and may be poorly matched to ReID. It proposes CDS, Top-k Sample Search, and FBLNeck to address architecture suitability, search efficiency, and loss incompatibility.
- Large classification backbones limit exploration of ReID-suitable architectures and require costly computation and inference time on resource-constrained devices.
- CDS uses CBlocks with two different kernel sizes to learn combined pattern information while keeping the search space GPU friendly.
- Top-k Sample Search computes only the highest-weight branches, reducing search cost while avoiding commitment to a single locally optimal branch.
- FBLNeck balances triplet and softmax losses by combining a Balance Neck with fine-grained local-feature extraction.
- CDNet achieves state-of-the-art performance on ReID and other tasks among lightweight networks.
2. Related Works
Related work spans compact CNNs, automated architecture search, and lightweight ReID models. Existing approaches reduce computation or add discriminative information, but many ReID systems still rely on costly backbones or parallel branches.
- MobileNets and ShuffleNets reduce parameters through depthwise separable or group convolution, while excessive branching can impede parallel computation.
- Many ReID models use ResNet backbones and extra structures for local, global, mask, camera, or view information, increasing deployment demands.
- Lightweight ReID research includes part-aware searched networks and OSNet, whose four parallel branches create heavy computational resource costs.
3. Methodology
The methodology defines CDS and its building blocks, searches layer-specific architectures with Top-k Sample Search, and adds FBLNeck to balance triplet and softmax supervision.
- 3.1. Combined Depth Space: CBlocks use two independent branches with different kernel sizes and fuse their heterogeneous features through channel-wise adaptive weighting.The two 1×1 convolutions squeeze and restore channels around the fusion operation.
- 3.1. Combined Depth Space: Each MBlock searches kernel combinations independently across layers, unlike DARTS, because different depths may focus on different information.The macro search network uses a stem, three stages, two MBlocks per stage, and fixed DownSample blocks between stages.
- 3.1. Combined Depth Space: CDS combines kernel choices with a depth factor, allowing the searched network depth to vary from 6 to 12 and be optimized rather than manually fixed.CDBlocks repeat CBlocks one or two times, expanding the search space to 12^6 ≈ 10^6.5 structures.
- 3.2. Top-k Sample Search: Top-k Sample Search computes only the highest-weight branches during forward propagation, then uses a differentiable bridge to update architecture parameters with their real weights.The selected branches are represented by a binary vector, while the gradient path uses the probability distribution without gradients for the discrete selection.
- 3.3. Fine-grained Balance Neck: FBLNeck contains a global BLNeck and a local FNeck that partitions an earlier feature map into two stripes to address inconsistent triplet and softmax targets.The neck inserts a fully connected layer between embeddings constrained by the two losses, and FNeck extracts local features from a relatively shallow layer.
- 3.2. Top-k Sample Search: The search alternates weight updates on training batches with architecture-parameter updates on validation batches, producing the final architecture from learned parameters.This avoids requiring convergence of both optimization problems at every iteration.
4. Experiments
Experiments evaluate the searched architectures, CDNet’s combined-pattern representations, and the FBLNeck across person ReID and classification settings. CDNet and its components show strong lightweight performance, while ablations support the roles of combined patterns and fine-grained features.
- Search for CNet and CDNet: Top-2 sample search produces the best architecture for both CNet and CDNet, whereas top-1 search can become trapped in a local architecture.Increasing the number of computational branches worsens performance because excess branches may compete.
- Evaluation on Person Re-Identification: CDNet and CNet surpass most published lightweight models across Market1501, DukeMTMC, and MSMT17 with few parameters.Table 3 covers models trained from scratch and ImageNet-pretrained models, with parameters counted at inference time.
- Visualization of Combined Pattern Learning: CDNet captures combined pedestrian patterns such as a skirt with a book, a handbag, and plain clothing patterns in its stage feature maps.The visualization uses the original image and the last feature maps from stages 1, 2, and 3.
- Classification Evaluation: 1.83% better classification performance than OSNet on CIFAR-100 supports the benefit of learning combined pattern information.The experiments compare CDNet with classification-oriented NAS methods and OSNet.
- Ablation Study: FBLNeck improves rank-1 accuracy and mAP by combining BLNeck with FNeck, while remaining removable at inference time.FNeck contributes fine-grained information from shallower depths, and FBLNeck extracts both local and global features.
- Ablation Study: Replacing CBlock’s combined branches with a single branch causes dramatic drops in rank-1 and mAP, supporting combined-pattern learning.The comparison uses CDNet std with the same number of building blocks.
5. Conclusion
The paper introduces CDS and uses top-2 sample search to obtain the lightweight CDNet for ReID. It also presents FBLNeck to balance triplet and softmax losses, and reports that CDNet outperforms state-of-the-art lightweight ReID networks.
- Conclusion: CDS yields the lightweight CDNet through top-2 sample search, while FBLNeck balances triplet-loss and softmax-loss effects during training.FBLNeck is described as effectively balancing the two losses, and CDNet is obtained from CDS using top-2 sample search.
7. Supplementary Material
Supplementary experiments examine searched architectures, feature-space behavior, model scaling, and transferability across ReID and ImageNet. CDNet remains computationally efficient while retaining competitive performance under reduced resources and on classification benchmarks.
- 7.1. Architecture Search Results: Top-2 sample search architectures almost achieve the best performance, while searched CDNet depths range from 6 to 12 without selecting the maximum depth.Top-2 CNet favors large-kernel combinations, compensating for its shallow depth.
- 7.2. Visualization of Feature Distribution: BLNeck maps triplet-loss features into an embedding space better suited to softmax loss, producing clearer angular margins between identities than BNNeck.The comparison is visualized through t-SNE projections of features before and after the neck.
- 7.3. Scaling CDNet with width multiplier and resolution multiplier: Reducing resolution from γ=1.0 to 0.5 lowers FLOPs significantly while rank-1 declines smoothly, whereas β=0.25 causes dramatic rank-1 and mAP drops.At β=0.25, stage-3 image resolution becomes 4×2, limiting effective feature learning.
- 7.3. Scaling CDNet with width multiplier and resolution multiplier: With 0.1M parameters and 77.9M FLOPs, CDNet achieves 91.7% rank-1 and 79.7% mAP when β=0.25 and γ=1.0.
- 7.5. Evaluation on ImageNet: On ImageNet, CDNet outperforms MobileNetV2 by 2.1% with fewer parameters and surpasses GDAS and DARTS by 1.1% and 1.8%, respectively.These results indicate transferability of the ReID-designed network to lightweight classification settings.
- 7.4. Comparison with State-of-the-Art Methods: With about 3× fewer FLOPs and 13× fewer parameters, CDNet achieves competitive performance and lower latency than BagofTrick, while outperforming two other lightweight models at faster speed.