RTSP Video Denoising Streamer

A real-time video streaming application that supports multiple denoising methods including ONNX models, PyTorch Lightning models, and OpenCV bilateral filtering.

Features

• Real-time video streaming with RTSP
• Multiple denoising methods:
  • ONNX models (CPU/GPU)
  • PyTorch Lightning models (CPU/GPU)
  • OpenCV bilateral filtering
• GPU acceleration support
• Configurable FPS and GPU memory usage
• Split-screen view of original and denoised video

Requirements

• Python 3.8+
• CUDA-capable GPU (optional, for GPU acceleration)
• FFmpeg installed on your system

Installation

1. Clone the repository:

git clone <repository-url>
cd denoise_vlc_rtsp

2. Create and activate a virtual environment (recommended):

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Install FFmpeg (if not already installed):

• Ubuntu/Debian:

sudo apt-get update
sudo apt-get install ffmpeg

• Windows: Download from FFmpeg website

Usage

Basic Usage

python utils/rtsp_streamer.py --method bilateral --video data/video/walking.mp4

Using ONNX Model

python utils/rtsp_streamer.py --method onnx --model path/to/model.onnx --video data/video/walking_noised.mp4

Using PyTorch Lightning Model

python utils/rtsp_streamer.py --method lightning --model models/lightning_model.ckpt --video data/video/walking.mp4

Advanced Options

python utils/rtsp_streamer.py \
    --method onnx \
    --model models/model.onnx \
    --video data/video/walking.mp4 \
    --fps 60 \
    --gpu-memory-fraction 0.8

Command Line Arguments

• --method: Denoising method to use (bilateral, onnx, or lightning)
• --model: Path to model file (required for onnx and lightning methods)
• --video: Path to input video file
• --fps: Target FPS for streaming (default: 60)
• --gpu-memory-fraction: Fraction of GPU memory to use (0.0 to 1.0, default: 0.8)

Viewing the Stream

1. Using VLC:

vlc udp://@127.0.0.1:1234

2. Using FFplay:

ffplay udp://127.0.0.1:1234

Performance Tips

1. For best GPU performance:

   • Use --gpu-memory-fraction 0.8 to leave some memory for system processes
   • Ensure your model is optimized for inference
   • Use appropriate batch sizes for your GPU memory

2. For CPU-only systems:

   • Use the bilateral filter method for better performance
   • Reduce the target FPS if needed

Troubleshooting

1. If you get CUDA out of memory errors:

   • Reduce the --gpu-memory-fraction
   • Use a smaller model
   • Reduce the input video resolution

2. If streaming is slow:

   • Check your network connection
   • Reduce the target FPS
   • Use a more efficient denoising method

3. If the video quality is poor:

   • Adjust the bilateral filter parameters
   • Use a higher quality model
   • Increase the bitrate in FFmpeg settings