Currently this repo will summarize and implement current loss functions and non-maximum suppression methods came up for object detection(simalar to the other repo obj-det-loss,but the other repo did not achieve a good result), so I work on this repo first.

Our goal is to analyze different tricks.

• Revise codes to be more readable and concise

• Loss_Funcs

  • bbox loss
    • Anchor-based Loss
      • YOLOv3-based
        • Regression Loss #testing
        • IOU Loss
        • GIOU Loss$[1]}#deal with gradient vanish caused by IOU is zero for non-overlap
        • Combined regression with GIOU
  • loss for confidence
    • Binary Cross Entropy
      • It is so hard to find a suitble pos/neg weight T T
    • dice loss[2]
      • hope to help deal with class imbalance
      • not so good as expect
  • Others
    • Use tanh to w,h to avoid grad explosion
    • Sum vs Mean
      • In my opinion, mean is literally better for analyze loss change but also means batch size and number of ground truth will influence gradient
      • even in some extent means errors in the crowded scenes(which is usually harder) get less penalty
      • So I display the mean(for analysis) but optimize on sum loss.

• Non-maximum-suppression

  • Hard NMS
  • Soft NMS[3]

• Other Tricks

  • Mosaic Augmentation
  • Sort ground truth to maximize matches reasonablly for multiple matches
  • Hard Key example mining and other machine learning tricks

• Results

  • On Validation Set
    • YOLO-SPP
      • YOLO Loss, mAP:0.63

• Reference:

  • [1]:"Generalized Intersection over Union: A Metric and A Loss for BOunding Box Regression":https://giou.stanford.edu/GIoU.pdf
  • [2]:"v-net loss"
  • [3]:"Soft-NMS -- Improving Object Detection With One Line of Code":https://arxiv.org/pdf/1704.04503.pdf