[論文レビュー] Bag of Tricks and A Strong Baseline for Deep Person Re-identification

Hao Luo, Youzhi Gu+3 more|arXiv (Cornell University)|Mar 17, 2019

Video Surveillance and Tracking Methods参考文献 27被引用数 121

ひとこと要約

この論文は、再識別のグローバル特徴量ベースラインを強力に構築するためのトレーニング tricks を集約・評価し、Market1501 で ResNet50 による最先端の結果を達成し、BNNeck やその他の改良を用いた DukeMTMC-reID でも競争力のある性能を示す。

ABSTRACT

This paper explores a simple and efficient baseline for person re-identification (ReID). Person re-identification (ReID) with deep neural networks has made progress and achieved high performance in recent years. However, many state-of-the-arts methods design complex network structure and concatenate multi-branch features. In the literature, some effective training tricks are briefly appeared in several papers or source codes. This paper will collect and evaluate these effective training tricks in person ReID. By combining these tricks together, the model achieves 94.5% rank-1 and 85.9% mAP on Market1501 with only using global features. Our codes and models are available at https://github.com/michuanhaohao/reid-strong-baseline.

研究の動機と目的

• Survey and consolidate effective training tricks used in person ReID.
• Build a strong, simple baseline based on global features without extra local-branch features.
• Quantitatively evaluate the impact of each trick on standard datasets.
• Provide a reference for fair comparison and practical baseline for industry applications.

提案手法

• Adopts a standard ResNet50 baseline with a global feature extractor.
• Introduces six tricks: warmup learning rate, Random Erasing Augmentation (REA), label smoothing, last-stride modification (stride=1), BNNeck, and center loss.
• Designs BNNeck to separate ID loss and triplet loss optimization by using a BN layer after feature extraction.
• Evaluates effects of tricks through ablation on Market1501 and DukeMTMC-reID, including cross-domain tests.
• Compares to state-of-the-art methods using only global features.
• Explores supplementary factors like batch size and image size on performance.

実験結果

リサーチクエスチョン

• RQ1 What is the performance gain from each training trick when added to a standard ReID baseline?
• RQ2 Can a strong baseline using only global features outperform methods that rely on multiple local features?
• RQ3 How do image size and batch size influence ReID performance, and how does BNNeck affect optimization dynamics?
• RQ4 Do training tricks transfer well across-domain (cross-domain) ReID scenarios?

主な発見

モデル	r = 1	mAP	r = 1	mAP
Baseline-S	87.7	74.0	79.7	63.7
+warmup	88.7	75.2	80.6	65.1
+REA	91.3	79.3	81.5	68.3
+LS	91.4	80.3	82.4	69.3
+stride=1	92.0	81.7	82.6	70.6
+BNNeck	94.1	85.7	86.2	75.9
+center loss	94.5	85.9	86.4	76.4

• A strong baseline with the six tricks achieves 94.5% rank-1 and 85.9% mAP on Market1501 using only global features.
• BNNeck provides the largest single-model improvement, especially on DukeMTMC-reID.
• Cross-domain results show warmup, label smoothing, and BNNeck significantly boost performance, while Random Erasing can hurt cross-domain performance.
• Compared to many state-of-the-art methods, global-feature baselines with these tricks can match or exceed performance, and with re-ranking (RK) can reach higher mAP and rank-1 scores.
• Increasing batch size and exploring image size have nuanced effects, with larger batch sizes generally helping hard-positive/negative mining and image size showing comparable performance across tested configurations.