CV-Lab3---Object-detection---Lego

Object detection using Faster R-CNN (ResNet-50)

Faster R-CNN (ResNet-50)

To run the Faster R-CNN (ResNet-50) model, please navigate to the Faster R-CNN folder and run the following command:

python3 main.py

This folder contains the natively trained Faster R-CNN (ResNet-50) model on the Lego dataset, with a fairly small data size (as of 31 Oct 2024), and equally small epoch, batch size, and learning rate.

All imgs and annotations were preprocessed and stored in the lego_dataset folder, re-labelled only in one "lego" class.

As for now, the result, mAP, is painful.

The YOLOV5 and Pytorch Approach

data_preprocessing.py

Use OpenCV for data loading and preprocessing

label_generator.py

To use yolov5 to train the model, the annotations should be converted from xml format to txt format, and the labels for each dataset are created under the label folder respectively.

Command to train the model

python train.py --data lego.yaml --weights yolov5s.pt --epochs 50 --img-size 640 --batch-size 16

Run the command under the yolov5 folder

Model summary: 157 layers, 7012822 parameters, 0 gradients, 15.8 GFLOPs Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 3/3 [00:18<00:00, 6.26s/it] all 75 252 0.949 0.966 0.989 0.895 Results of training the model are saved to /yolov5/runs/train/exp3

To view the mAP results after training

python val.py --data lego.yaml --weights runs/train/exp/weights/best.pt --img-size 640 --iou-thres 0.5

Run the following command to see how well the model performs on new images:

python detect.py --weights runs/train/exp/weights/best.pt --source /path/to/test/images --img-size 640 --conf-thres 0.5

Tianyu-Fang/CV-Lab3---Object-detection---Lego