Stroma CV Challange (Multi-class Object Detection and Counting)

Processing Units

• NVIDIA A100-SXM4-40GB, 40536MiB (google colab, training)
• GTX 1050TI for (local pc, inference)

Changed Hyperparameters

• Batch sized increased to 16, lr selected as 0.01
• Confidence threshold increased to 0.65 to eliminate false detections
• trained yolov5 and yolov8 models and achieved best overall accuracy was visually %100 on test video with yolov5
• albumentations (horizontal flip, jitter, scale)

Added Features

• Track labels contains class names and track trace line removed

Pre-Requsities

• Python 3.9 (Python 3.7/3.8 can work in some cases)
• pip install -r requirements.txt

data download

• images, labels, custom yml file and trained model "best.pt" (yolov5)
• https://drive.google.com/drive/folders/1_h87v1ts_kUqZ5n5fOfQhJkevCvS-LdA?usp=sharing

for object detection + object tracking + labels(edited v2 version)

• python obj_det_and_trk_2.py --weights yolov5s.pt --source "your video.mp4"

What is next (working on)

• ✨ Extra: Inference optimizations (e.g. pruning, quantization) with libraries like TensorRT, ONNX runtime and a comparison of each
• ✨ Extra: Optimizing inference pipeline to work on an embedded system like a Jetson in real time. (>30fps)

References

• https://github.com/ultralytics/yolov5
• https://github.com/RizwanMunawar/yolov5-object-tracking
• https://github.com/abewley/sort