[논문 리뷰] Margin Sample Mining Loss: A Deep Learning Based Method for Person Re-identification

Qiqi Xiao, Hao Luo외 1명|arXiv (Cornell University)|2017. 10. 02.

Video Surveillance and Tracking Methods참고 문헌 44인용 수 131

한 줄 요약

MSML는 배치 내에서 가장 어려운 양의 페어와 가장 어려운 음의 페어를 선택하여 임베딩 모델을 학습하는 하드 샘플 경계 기반의 메트릭 학습 손실을 도입하고, 주요 벤치마크에서 여러 최첨단 손실들을 능가합니다.

ABSTRACT

Person re-identification (ReID) is an important task in computer vision. Recently, deep learning with a metric learning loss has become a common framework for ReID. In this paper, we also propose a new metric learning loss with hard sample mining called margin smaple mining loss (MSML) which can achieve better accuracy compared with other metric learning losses, such as triplet loss. In experi- ments, our proposed methods outperforms most of the state-of-the-art algorithms on Market1501, MARS, CUHK03 and CUHK-SYSU.

연구 동기 및 목표

• Motivate improved metric learning for person ReID beyond standard triplet/quadruplet losses.
• Propose a margin sample mining loss that exploits extremely hard samples within a batch.
• Demonstrate that MSML yields superior performance on major ReID benchmarks.

제안 방법

• Construct a batch with K images per identity (P identities, N=K×P).
• Compute an N×N distance matrix over embedding features from a CNN backbone.
• Select the hardest positive pair (maximum distance among positive pairs) and the hardest negative pair (minimum distance among negative pairs).
• Define the MSML loss as L_eml = (max_positive_distance − min_negative_distance + α)+, where (·)+ is hinge.
• Incorporate edge mining by handling cases where positive and negative pairs share identities and cases where they do not.
• Train with Adam using staged learning rates and standard data augmentation.

실험 결과

연구 질문

• RQ1 Does MSML improve discrimination between same-identity and different-identity pairs compared to existing metric losses?
• RQ2 How does hard/edge mining in MSML affect performance across different backbone networks and datasets?
• RQ3 Can MSML achieve state-of-the-art results on Market1501, MARS, CUHK-SYSU, and CUHK03?
• RQ4 What is the impact of MSML on the embedding space structure (distance distributions) compared to triplet/quadruplet losses?

주요 결과

Dataset	Model/Loss	mAP	r=1	r=5	r=10
Market1501	MSML (ResNet50)	69.6	85.2	93.7	-
Market1501	MSML (Inception-v2)	73.4	87.7	95.2	-
Market1501	MSML (ResNet50-X)	76.7	88.9	95.6	-
MARS	MSML (ResNet50)	72.0	83.0	92.6	-
MARS	MSML (Inception-v2)	74.6	84.2	95.1	-
MARS	MSML (ResNet50-X)	72.0	83.4	93.3	-
CUHK-SYSU	MSML (ResNet50)	87.2	89.3	96.4	-
CUHK-SYSU	MSML (Inception-v2)	88.4	90.4	96.8	-
CUHK-SYSU	MSML (ResNet50-X)	89.6	90.9	97.4	-
CUHK03	MSML (ResNet50)	84.0	96.7	98.2	-

• MSML achieves best accuracy on most experiments across base models and datasets tested.
• On Market1501 with ResNet50, MSML yields mAP 69.6 and rank-1 85.2; on MARS, mAP 72.0 and rank-1 83.0; on CUHK-SYSU, mAP 87.2 and rank-1 89.3; on CUHK03, mAP 84.0 and rank-1 96.7.
• Compared to TriHard and Quad, MSML consistently offers superior or competitive results across configurations.
• Distance distribution analyses show MSML yields a clearer separation between positive and negative pairs than traditional triplet-based losses.