Bag of Tricks and A Strong Baseline for Deep Person Re-identification

Luo, Hao; Gu, Youzhi; Liao, Xingyu; Lai, Shupeng; Jiang, Wei

doi:10.48550/arxiv.1903.07071

Cited by 9 publications

(7 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…a) original image b) EGCSA c)bagtricksOur attention module effectively identifies distinct features of each individual, separating them from the background with clarity. This demonstrates the efficiency of our proposed attention modules compared to BagTricks[4], as our method identifies more discriminative features of persons in the images.…”

mentioning

confidence: 76%

“…On the DukeMTMC-ReID [32] dataset, compared to the baseline (BagTricks [4]), we achieved an eight percent improvement in mAP. our smaller EGCSA-4 model yielded comparable results to other models while maintaining a smaller size.…”

Section: Comparison To State-of-the-art Methodsmentioning

confidence: 95%

“…They used GiLt for training part-based representation. To enhance similarity metrics, some methods [4,14,21,43] combine triplet loss with identification loss and jointly learn a metric within the model. Zhou et al [44]propose a Sparse Pairwise (SP) loss that only uses a few appropriate pairs for each class in a minibatch.…”

Section: Related Work 31mentioning

confidence: 99%

“…1 in real-world scenarios can still impact the accuracy of person re-identification. Previous studies [4,5] have shown that using only convolutional neural networks can make it difficult to distinguish individuals with similar appearances or those obscured by other people or objects.…”

Section: Introductionmentioning

confidence: 99%

“…Using BagTricks [4], we have designed a network that embeds EGCSA into various layers. This approach allows us to capture more discriminative features, enabling us to effectively address issues like occlusion and similar appearances which are very common in Person Re-id.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Person Re-Identification Across Multiple Cameras: An Efficient Gated Channel-Spatial Attention Method

golizadeh

2024

Preprint

View full text Add to dashboard Cite

Person re-identification is a challenging task due to various factors such as body position variations, lighting changes, visual obstructions, and complex backgrounds. In this paper we introduce an attention module called efficient gated channel-spatial attention (EGCSA) to address these challenges. EGCSA consists of two submodules: efficient channel-wise attention (ECAt) and efficient spatial attention (ESAt). ECAt focuses on interactions at the channel level within feature maps by generating attention weights through convolutional layers. ESAt operates at the spatial level to capture relevant information across different spatial locations. These submodules produce attention weights. A multi-stage gate mechanism selects efficient features from the generated ones in previous step, enabling the network to learn robust and discriminative features. The proposed module is integrated into the backbone network after multiple layers. By leveraging EGCSA, our approach demonstrates improved performance in person re-identification tasks, showcasing the effectiveness of our attention mechanism in handling challenging re-identification scenarios. Experimental results on Market-1501, DukeMTMC-ReID, and MSMT17 datasets show that our method consistently outperforms existing state-of-the-art approaches. Code is available at https://github.com/Mahdi-Golizadeh/EGCSA-person-reid.git.

show abstract

mentioning

confidence: 76%

Section: Comparison To State-of-the-art Methodsmentioning

confidence: 95%