Open xujiayi-hub opened 1 month ago
BCE loss exhibits inferior performance due to imbalance between positive and negative samples. Following ACM, ALCNet, DNANet, we employ soft-IoU loss. More discussion can be seen in 《Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision》.
In most referring papers for example UIUNet, they use BCE loss for optimization. Why you choose to use soft-IoU loss in most models?