tagomaru/yolo-heatmaps

A utility for generating heatmaps of YOLOv5 using Layerwise Relevance Propagation (LRP/CRP).

YOLOv5 Heatmaps

A utility for generating heatmaps of YOLOv5 https://github.com/ultralytics/yolov5 using Layerwise Relevance Propagation (LRP/CRP). Pytorch implementation based on: https://github.com/moboehle/Pytorch-LRP

Install

git clone https://github.com/akarasman/yolo-heatmaps/
cd yolo-heatmaps
pip install -r requirements.txt

CLI Use Example

python3 explain.py --source=data/images/so-and-so.jpg --weights=yolov5s.pt --explain-class='person'

Run results saved to runs/explain/exp(# of run)

Arguments

  # explain.py is built on detect.py module from YOLOv5, lrp options are :
  
  --power POWER         Power exponent applied to weights and inputs
  --contrastive         Use contrastive relevance (CRP)
  --b1 B1               Visualization parameter for CRP - multiplier of primal part
  --b2 B2               Visualization parameter for CRP - multiplier of dual part
  --explain-class EXPLAIN_CLASS
                        Class to explain
  --conf                Confidence threshold on object
  --max-class-only      Max class only
  --box-xywh BOX_XYWH [BOX_XYWH ...]
                        Box to restrict investigation (X,Y,W,H format)
  --smooth-ks SMOOTH_KS
                        Box to restrict investigation (X,Y,W,H format)
  --box-xyxy BOX_XYXY [BOX_XYXY ...]
                        Box to restrict investigation (X,Y,X,Y format)
  --cmap CMAP           Explanation color map (default set to seismic/magma when contrastive / non-contrastive

Current version only supports YOLOv5s-x models.

Please cite our paper if you plan on using code from this repository for your work

@inproceedings{inproceedings,
author = {Karasmanoglou, Apostolos and Antonakakis, Marios and Zervakis, Michalis},
year = {2022},
month = {06},
pages = {1-6},
title = {Heatmap-based Explanation of YOLOv5 Object Detection with Layer-wise Relevance Propagation},
doi = {10.1109/IST55454.2022.9827744}
}