Object Detection - yolov10

installation

This project use uv to manage Python version, pakcage dependencies and virtual environment.
Yes, uv is the replacement of pyenv, pip and venv.

install uv

# Install pipx through pip
pip install pipx

# Install uv through pipx
pipx install uv

So, what is the next? No, you are ready to run server/client.

uv will automatically create venv, donwload packages when you run uv run SCRIPT.py. And, that is blazing fast.

run server

make server
# or in development
make dev

test by shell

It will read and send one local jpg file to our server to get the result of object detection.

curl -X 'POST' \
  'http://127.0.0.1:8000/predict/' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'file=@./data/office_parkour1.jpg' -o detect.json

test by client.py

It will get all the frames in given video file (`data/Parkour...mp4 there), and send them to server to do inference one by one. Then, it will open cv2 to stream these frames were detection by YOLOv10.

I only cut ~13 seconds out from the original video. About 400 frames under 30 FPS.

make client

Try to speedup

Start from FPS around 18.

Some optimizations:

• try differnt compression package to minimize the size of jpg (use simplejpeg there)
• disable default verbose in server
• add stream=True in server
• add torch.comple Did not applied because my GPU of development environment is too old.
  RuntimeError: Found NVIDIA GeForce GTX 1060 6GB which is too old to be supported by the triton GPU compiler, which is used as the backend.
  Could try another GPU card, colab or Cloud platform.

Improve to around 21

• only return information of bounding boxes instead the whole jpg. Improve to around 25 FPS
• truncate the length of message (about half) Improve to around 28 FPS

There is a cap of 30 FPS because that is the FPS of the original video.

At the final setting, YOLOv10m runs with 21 FPS, and YOLOv10l go with 17 FPS.