• 调整UFDN模型通道划分的约束损失的权重进行实验。
• 对局部特征进行多尺度空间池化，再融合特征。
• 考虑到模型局部的id损失可能不太准去，想到使用全局对局部进行自蒸馏的结构。
• Tim回复相关工作。

去掉通道约束后结性能反而提升

2023-05-03 08:51:30,270 reid_baseline INFO: Global Results
2023-05-03 08:51:30,270 reid_baseline INFO: mAP: 79.7%
2023-05-03 08:51:30,270 reid_baseline INFO: mINP: 42.9%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-1  :96.6%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-5  :98.5%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-10 :99.0%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-100:99.8%
2023-05-03 08:51:30,270 reid_baseline INFO: decouple Merged Results
2023-05-03 08:51:30,270 reid_baseline INFO: mAP_2: 82.4%
2023-05-03 08:51:30,270 reid_baseline INFO: mINP_2: 47.1%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-1  :97.0%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-5  :98.6%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-10 :99.2%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-100:99.9%
2023-05-03 08:51:30,270 reid_baseline INFO: All Merged Results
2023-05-03 08:51:30,270 reid_baseline INFO: mAP_3: 82.4%
2023-05-03 08:51:30,270 reid_baseline INFO: mINP_3: 47.2%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-1  :97.0%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-5  :98.6%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-10 :99.3%
2023-05-03 08:51:30,270 reid_baseline INFO: CMC curve, Rank-100:99.9%

去掉通道约束后结性能反而提升

2023-05-06 09:54:12,931 reid_baseline INFO: Global Results
2023-05-06 09:54:12,932 reid_baseline INFO: mAP: 82.1%
2023-05-06 09:54:12,932 reid_baseline INFO: mINP: 47.0%
2023-05-06 09:54:12,932 reid_baseline INFO: CMC curve, Rank-1  :97.5%
2023-05-06 09:54:12,932 reid_baseline INFO: CMC curve, Rank-5  :98.6%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-10 :98.9%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-100:99.9%
2023-05-06 09:54:12,933 reid_baseline INFO: decouple Merged Results
2023-05-06 09:54:12,933 reid_baseline INFO: mAP_2: 83.2%
2023-05-06 09:54:12,933 reid_baseline INFO: mINP_2: 49.2%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-1  :97.6%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-5  :98.7%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-10 :99.2%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-100:99.9%
2023-05-06 09:54:12,933 reid_baseline INFO: All Merged Results
2023-05-06 09:54:12,933 reid_baseline INFO: mAP_3: 83.3%
2023-05-06 09:54:12,933 reid_baseline INFO: mINP_3: 49.4%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-1  :97.5%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-5  :98.7%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-10 :99.2%
2023-05-06 09:54:12,933 reid_baseline INFO: CMC curve, Rank-100:99.9%

问题分析

• 去掉通道约束的损失后，性能反而增加
  之前整体模型性能是82.0。在消融实验时，需去掉通道划分局部的约束损失进行训练。结果发现，整体模型去掉对于通道的约束后，性能反而更好，达到了82.4。反复实验，并对损失进行权重调整后，确定了其无法让模型性能更高，无法作为创新点使用。
  因此只能想新的创新点，最开始考虑在对局部特征空间注意力后的GAP进行优化，将其修改为多尺度金字塔池化，然后再把特征加在一起。实验结果一般，并且考虑到这也许只能和空间注意力一起作为一个创新点，还是没法写论文。
  在UFDN模型中，每个局部都使用id损失，考虑到对局部特征进行id损失可能没有全局那么准确，因此受到自蒸馏模型的启发，使用全局特征对局部特征进行蒸馏。最终性能最高达到83.5。