level
B: 开源了code - /kding1225/TDTS-visdrone,ablation study 也还算扎实,等我跑一遍看看
intro
跑的 VisDrone-Det ,对比的是FCOS,精度很低,对应的ICCV workshop官方论文里给的top1是29+,这篇里面的精度是23,对应的top1 找到一篇报导 大致是cascade+dcn+no-local+ensemble,不过现在官网的leadboard给的精度非常高 /leaderboard/ 不太懂为什么和ICCVW 里面差距很大,莫非是后续可以提交,希望道友评论区告知
有意思的点
提到的文献:
Small object detection in unmanned aerial vehicle images using feature fusion and scaling-based single shot detector with spatial context analysis | 引用也有个十几SCRDet ++ : Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing | yangxue大佬的Convolutional SVM Networks for Object Detection in UAV Imagery | citation 64
疑惑
这个prophet head 按理是用来在forward阶段进行区域筛选的,看起来没有这个操作啊
bn换成这个cn真的有用吗,看最后的精度涨点是千分位的变动,很可能不work
创新
column normalization
We thus propose an alternative, column normalization (CN), which only computes statistics along the channel dimension
In other words, the features are regarded to be independent of batch, width, and height. Once mean and variance are computed, and learnable linear transform is applied to each feature plane
采用了sparse conv
Better implementation of sparse convolution must lead to higher speedups against the vanilla convolution
采用了 focal loss
We adopt focal loss [27] as the positive and negative samples are imbalanced