ReVo

<div align="center">

ReVo: A Cross-Layer Reliable Volumetric Videoconferencing System

Ankur Aditya^1*, Diptyaroop Maji^1*, Lingdong Wang¹, Bhavya Ramakrishna², Ramesh Sitaraman^1,3, Prashant Shenoy¹

¹University of Massachusetts Amherst     ²Dolby Labs     ³Akamai Tech <br> ^* Student authors with equal contribution

</div>

This repository contains the pre-trained neural loss recovery module checkpoints for the ReVo system. These models are designed to perform neural loss recovery for volumetric videoconferencing systems, recovering lost packets across different video codecs.

Model Details

The repository includes 6 distinct models, organized by the compression codec and the RGB and Depth modality.

Supported Codecs:

• h264 (Advanced Video Coding)
• h265 (High Efficiency Video Coding / HEVC)
• dcvcrt (Deep Contextual Video Compression)

Repository Structure

umass-lass/ReVo
├── h264/
│   ├── h264_rgb.pth
│   ├── h264_depth.pth
├── h265/
│   ├── h265_rgb.pth
│   ├── h265_depth.pth
└── dcvcrt/
    ├── dcvcrt_rgb.pth
    ├── dcvcrt_depth.pth

How to Use?

from huggingface_hub import hf_hub_download
import torch

# Define the specific codec and modality you want to load
codec = "h265"      # Options: 'h264', 'h265', 'dcvcrt'
modality = "depth"  # Options: 'rgb', 'depth'

# Download the specific checkpoint
checkpoint_path = hf_hub_download(
    repo_id="umass-lass/ReVo", 
    filename=f"{codec}/{codec}_{modality}.pth"
)

# Load the weights into your PyTorch model
# model.load_state_dict(torch.load(checkpoint_path))
print(f"Successfully downloaded to: {checkpoint_path}")

Citation

(Coming Soon)