论文阅读
Multi-ReceptiveField Soft Attention Part Learning for Vehicle Reidentification

代码工作

• 使用Orientation-aware论文中提到的diverse损失函数约束空间注意力。
• 使用Multi-Receptive Field Soft Attention Part Learning 论文中的diverse损失函约束空间注意力。
• 修改代码使用使用原始不分割的特征计算出代表背景信息的注意力，最后使用softmax来产生注意力图。
• 使用nn.sigmoid 生成注意力图，并用softmax对所生成的注意力进一步处理。
• 修改diverse 损失函数的权重进一步进行实验。

2023-04-11 03:43:47,761 reid_baseline INFO: Global Results
2023-04-11 03:43:47,761 reid_baseline INFO: mAP: 80.7%
2023-04-11 03:43:47,761 reid_baseline INFO: mINP: 43.0%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-1  :96.5%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-5  :98.3%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-10 :99.0%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-100:99.9%
2023-04-11 03:43:47,761 reid_baseline INFO: decouple Merged Results
2023-04-11 03:43:47,761 reid_baseline INFO: mAP_2: 81.5%
2023-04-11 03:43:47,761 reid_baseline INFO: mINP_2: 48.0%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-1  :96.7%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-5  :98.5%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-10 :99.2%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-100:99.9%
2023-04-11 03:43:47,761 reid_baseline INFO: All Merged Results
2023-04-11 03:43:47,761 reid_baseline INFO: mAP_3: 81.5%
2023-04-11 03:43:47,761 reid_baseline INFO: mINP_3: 48.1%
2023-04-11 03:43:47,761 reid_baseline INFO: CMC curve, Rank-1  :96.5%
2023-04-11 03:43:47,762 reid_baseline INFO: CMC curve, Rank-5  :98.3%
2023-04-11 03:43:47,762 reid_baseline INFO: CMC curve, Rank-10 :99.2%
2023-04-11 03:43:47,762 reid_baseline INFO: CMC curve, Rank-100:99.9%

2023-04-12 05:16:47,447 reid_baseline INFO: mAP: 79.8%
2023-04-12 05:16:47,447 reid_baseline INFO: mINP: 43.0%
2023-04-12 05:16:47,447 reid_baseline INFO: CMC curve, Rank-1  :96.2%
2023-04-12 05:16:47,447 reid_baseline INFO: CMC curve, Rank-5  :98.7%
2023-04-12 05:16:47,447 reid_baseline INFO: CMC curve, Rank-10 :99.1%
2023-04-12 05:16:47,447 reid_baseline INFO: CMC curve, Rank-100:99.8%
2023-04-12 05:16:47,447 reid_baseline INFO: decouple Merged Results
2023-04-12 05:16:47,447 reid_baseline INFO: mAP_2: 81.4%
2023-04-12 05:16:47,447 reid_baseline INFO: mINP_2: 47.3%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-1  :97.1%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-5  :99.0%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-10 :99.4%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-100:99.9%
2023-04-12 05:16:47,448 reid_baseline INFO: All Merged Results
2023-04-12 05:16:47,448 reid_baseline INFO: mAP_3: 81.4%
2023-04-12 05:16:47,448 reid_baseline INFO: mINP_3: 47.3%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-1  :97.0%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-5  :99.0%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-10 :99.5%
2023-04-12 05:16:47,448 reid_baseline INFO: CMC curve, Rank-100:99.9%

问题分析

• 使用 diverse 损失后性能变差
  总结目前的实验结果，不使用diverse 损失时，模型性能可以达到81.7。添加后diverse后性能下降到79，接着在nn.sigmoid 后进一步使用 softmax，性能提升到81.5.以及一些其他对diverse 损失函数的修改以及损失权重的调整，性能也没有比81.5 更好。

下一步

总结目前的实验结果，分析问题，进一步调整参数进行实验。