KananVyas/person_car_detection_yolov5

Person and Car Detector

Person and Car Detection

This repo contains transfer learning approach for fine-tune person and car detection problem.

Training Script:

In train.ipynb

• We have used YOLOv5 to train the current model.

• The data preprocessing includes conversion of COCO labels to YOLO txt format.

• The training has been done on 2000 images and validated on 239 samples for 10 epochs.

• We have used yolov5m.pt as our pretrained checkpoint. Alternatively, we can use higher accurate heavy checkpoints from here: https://github.com/ultralytics/yolov5#pretrained-checkpoints

• The training has been done on google colab, checkout the code using colab icon below:

Accuracy Metrics

With this training, mAP@0.5 achieved 71% accuracy on validation set.

• As seen above, all classes are captured with mAP@0.5=0.71 at the threshold of >0.3.

• Hence, we can set our confidence threshold between 0.25-0.3.

Testing Script

with test.ipynb

• We can load the trained model from the given path.
• Infer the model and plot the detections using the script.
• Run this script with google colab:

Docker support and API:

Using /api/main.py:

• We can run inference using FastAPI endpoint.
• Execute python3 /api/main.py to run uvicorn server on port 7000. It will automatically load the model and by calling /predict endpoint in Postman, we can get inference JSON as an output.
• To run in containarized env, build the Dockerfile
• • docker build -f ./api/Dockerfile -t person_car_api .
• After succesfully building the docker,
• • docker run -it --net host person_car_api
• After running the container, call http://127.0.0.1:7000/predict POST request and send image as file form-data to the endpoint.
• The output json structure will look like:

{"output":  [{
	"xmin":  14.2618331909,
	"ymin":  76.5265274048,
	"xmax":  380.9652099609,
	"ymax":  397.6776123047,
	"confidence":  0.8891581893,
	"class":  0,
	"name":  "person"
	}],
"inference_time":  556.9939613342285 }