Combine YOLOv3 with MiDaS with a single Resnext101 backbone. The idea is to use a single feature extractor for two different applications, in this case, Object detection and Monocular Depth Estimation.
Please read the article to see more details https://sarvan0506.medium.com/yolo-v3-and-midas-from-a-single-resnext101-backbone-8ba42948bf65
The model architecture change can be seen in model/mde_net.py
The model is trained on Construction Safety Gear Data which can be found here https://github.com/sarvan0506/EVA5-Vision-Squad/tree/Saravana/14_Construction_Safety_Dataset. If training need to done on custom datasets refer the data preparation steps mentioned in the page.
Place the data inside data/customdata/custom.data folder
python3.6 train.py --data data/customdata/custom.data --batch 8 --cache --cfg cfg/mde.conf --epochs 50 --img-size 512
Please refer the config file cfg/mde.cfg to change the network configuration, freeze different branches. The model is an extension of YOLOv3 and MiDaS networks. Most of the configurations can be understood if familiar with
Download the weights from https://drive.google.com/file/d/1LZoWaZbsD4gG4xgWQ4cW-ezyhmaHXV1O/view?usp=sharing and place it under weights folder
Place the images on which inference need to be run, inside input folder
python3.6 detect.py --source input --conf-thres 0.1 --output output --weights weights/best.pt
The inferred images will be stored inside output folder
