Object detection loss functions and non-maximum suppressions Survey

This repo will summarize and implement current loss functions and non-maximum suppression methods came up for object detection.

All methods will be evaluated on VOC2012

• Ref:https://github.com/eriklindernoren/PyTorch-YOLOv3
  • Thanks eriklindernoren's share, it's really save me much time while debuging
  • models.py and parse_config and some other function are copied from his repo to help me check my reproduce correctness
  • you can use cmp_network parameter and add sth like this to Loss API to check outputs/structure of the network and the loss
• A little tip:
  • watch out every hyperparameter while reporducing codes from scratch, even a little difference can make you spending half of a year to debug:(
• TBH, still working on hyperparameters tuning, but the whole code should work fine, someday.

Template Code Structure for Deep Learning

• Data
  • dataset annotation
  • dataset pre-processing/analyze
• Models
  • utils - necessary function
    • maybe move evaluation metric here?
  • network
  • loss
  • backbone network
  • network configuration
• dataProcessing
  • dataloader for train and test
• Utils
  • logger
  • evaluation/metric
  • non-maximum-supression
• visualization - jupyter notebook
• Trainer - APIs
  • save/load weights
  • lr scheduler
  • optimizer(sometimes need to pass in network for things like GAN)
• train
• test
• evaluation - evaluate predicted result
• config - configurate parpameters

To do List

• Revise codes to be more readable and concise
• Loss_Funcs
  • bbox loss
    • Anchor-based Loss
      • YOLOv3-based
        • Regression Loss
        • IOU Loss
        • GIOU Loss$^{[1]}$#deal with gradient vanish caused by IOU is zero for non-overlap
        • Combined regression with GIOU
  • classification loss
    • [x]dice loss$^{[2]}$ #help deal with class imbalance
  • Anaylysis
  • Innovation
• Non-maximum-suppression
  • Hard NMS
  • Soft NMS$^{[3]}$

[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