yowov3-multistreaming

YOWOv3(Spatio Temporal Action Detection task) using (UCF101-24) dataset. The repo is extension of https://github.com/Hope1337/YOWOv3, https://arxiv.org/pdf/2408.02623

Environment Setup

Clone this repository:

git clone https://github.com/irfan112/yowov3-multistreaming-inferencing.git

Use Python 3.8 or Python 3.9, and then install the dependencies:

pip install -r requirements.txt

pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 \
  --extra-index-url https://download.pytorch.org/whl/cu117

Datasets

UCF101-24

Download from: Google Drive Link

Pretrained Weights & Checkpoints

To train or evaluate YOWO (I3D / ResNet), you need to download the pretrained weights and checkpoints provided here:

Google Drive - YOWO (I3D / ResNet) Checkpoints

After downloading, place the files into the corresponding weights/ or checkpoints/ folder in this repository (create them if they don’t exist).

yowov3-multistreaming-inferencing/
│── weights/
│   ├── yowo_i3d.pth
│   ├── yowo_resnet.pth
│── checkpoints/
│   ├── checkpoint_epoch_XX.pth

Video Source Configuration

You can specify your video inputs in a .env file. These sources can be either local video files or RTSP streams from live cameras.

Example (.env)

# Primary video sources (can be RTSP streams or video files)
VIDEO_SOURCE_1=ucf24/videos/Basketball/v_Basketball_g22_c01.mp4

# Example RTSP live camera stream
VIDEO_SOURCE_1=rtsp://admin:password@192.168.1.100:554/cam/realmonitor?channel=1&subtype=1

Once your .env is configured, run YOWOv3 with one of the following modes:

🔹 Multistreaming Mode

python main.py -m multistreaming_live -cf config/cf2/ucf_config.yaml

🔹 Single Live Mode (Original YOWOv3)

python main.py -m live -cf config/cf2/ucf_config.yaml

Multistreaming Example

Basketball Stream

Diving Stream

Training Info – YOWO (I3D & ResNet-3D Backbones)

trained YOWO using two different 3D backbones: I3D and ResNet-3D.

🔹 Observations

• I3D Backbone started with a lower loss and converged more smoothly.
• ResNet-3D Backbone had a higher initial loss but showed consistent improvement and comparable convergence by epoch 7.
• Both models benefited from gradual learning rate decay.

References

• Extension of Hope1337 / YOWOv3
• YOWOv2 implementation — yjh0410 / YOWOv2
• Related document / paper overview — “YOWOv3: A Lightweight Framework for Spatio-Temporal Action Detection” on Studocu
  (https://www.studocu.vn/vn/document/van-lang-university/tu-tuong-ho-chi-minh/yowov3-a-lightweight-framework-for-spatio-temporal-action-detection/130031369)