Experiments

based on new-yolov1-pytorch project 2.4 IoU Loss IoU Loss Model mAP(07test) LogFile FPN/MultiHeadFPN 0.7149 eval_log/log_eval_myYOLOWithFPNMultiPred_with_sam_for_3_head_142 IoULoss replace origin txtytwth loss 0.558 log_myYOLOWithFPNMultiPredWithIoULoss_iouweight10_yolo_160 添加中心点距离最近的target assign机制，避免按label顺序匹配gt object对应的特征点，IoULoss replace origin txtytwth loss 0.571 log_myYOLOWithFPNMultiPredWithIoULoss_iouweight10_targetassian_by_min_dist_yolo_160 GIoULoss Model mAP(07test) LogFile FPN/MultiHeadFPN 0.7149 eval_log/log_eval_myYOLOWithFPNMultiPred_with_sam_for_3_head_142 GIoULoss replace origin txtytwth loss 0.6674 log_myYOLOWithFPNMultiPredWithGIoULoss_SGD_iouweight1_tvgiouloss_sum_target_assign_by_min_dist_yolo_130 DIoULoss CIoULoss SIoULoss EIoULoss Luxury IoU Loss: condition mAP txtytwth_iou_weightsum_loss = txtytwth_loss + giou_loss * 1

based on new-yolov1-pytorch project 2.1 FocalLoss: PaperMode blog MultiHeadFPN+FocalLoss Model mAP(07test) LogFile FPN/MultiHeadFPN 0.7149 eval_log/log_eval_myYOLOWithFPNMultiPred_with_sam_for_3_head_142 Loss/MultiHeadFPNFocalloss alpha=0.75 0.6742 eval_log/ Loss/MultiHeadFPNFocalloss alpha=0.5 0.6964 eval_log/log_myYOLOWithFPNMultiPredFocalLoss_yolo_143_new Loss/MultiHeadFPNFocalloss alpha=0.25 0.7121 eval_log/log_myYOLOWithFPNMultiPredFocalLoss_alpha0.25_yolo_154 BiFPN+FocalLoss Model mAP(07test) LogFile FPN/MultiHeadFPN 0.7149 eval_log/log_eval_myYOLOWithFPNMultiPred_with_sam_for_3_head_142 Loss/MultiHeadBiFPNFocalLoss alpha=0.5 0.6975 eval_log/log_myYOLOWithBiFPNMultiPredFocalLoss_yolo_152 Loss/MultiHeadBiFPNFocalLoss alpha=0.25 0.7201 eval_log/log_myYOLOWithBiFPNMultiPredFocalLoss_alpha0.25_yolo_141 AugFPN+FocalLoss Model mAP(07test) LogFile FPN/MultiHeadFPN 0.7149 eval_log/log_eval_myYOLOWithFPNMultiPred_with_sam_for_3_head_142 Loss/MultiHeadAugFPNFocalLoss alpha=0.5 0.6922 eval_log/log_myYOLOWithAugFPNMultiPredFocalLoss_yolo_143 Loss/MultiHeadAugFPNFocalLoss alpha=0.25 0.7128 eval_log/log_myYOLOWithAugFPNMultiPredFocalLoss_alpha0.25_yolo_154 2.2 PolyFocalLoss Model mAP(07test) LogFile FPN/MultiHeadFPN 0.7149 eval_log/log_eval_myYOLOWithFPNMultiPred_with_sam_for_3_head_142 Loss/MultiHeadFPNPolyLossFL poly_scale=1 poly_pow=1 0.7182 Loss/MultiHeadFPNPolyLossFL poly_scale=1 poly_pow=2 0....

目录： BCE与CE公式的差别 Entropy Cross Entropy 二分类交叉熵损失 Binary Cross Entropy 如果将此思路扩展到多分类 公式解释 多分类交叉熵损失 Multi-classes Cross Entropy 标准的多分类交叉熵损失 注意事项 总结 Focal Loss理解 代码实现两种CE+FocalLoss 设计实验 BCE与CE公式的差别 Entropy 一个分布中的信息熵： $$ H(p) = - \sum_{i} p_i \log(p_i) $$ Cross Entropy 两个分布的交叉熵： $$ \text{Cross-Entropy} = - \sum_{i} p_i \log(q_i) $$ 二分类交叉熵损失 Binary Cross Entropy 在二分类问题中，交叉熵损失函数同时考虑了正类和负类的预测损失。公式如下： $$ \text{Binary Cross-Entropy Loss}(y, \hat{y}) = - \left[ y \log(\hat{y}) + (1 - y) \log(1 - \hat{y}) \right] $$ 其中，$(1 - y) \log(1 - \hat{y})$ 是对负类（错误分类）的惩罚。...