SceneDiff Annotator

A video annotation tool for logging down changes between paired video sequences.

🔗 Check out the project page for more details.

Overview

The SceneDiff Annotator is built on top of SAM2 and provides a complete workflow for annotating changes between video pairs:

• Upload & Configure: Upload video pairs and specify object attributes (deformability, change type, multiplicity)
• Interactive Annotation: Provide sparse point prompts on selected frames with an intuitive click-based interface
• Offline Propagation: Automatically propagate masks throughout both videos
• Review & Refine: Visualize annotated videos, refine annotations, and verify results

Demo

annotation_visualization.mov

Watch the video above to see the annotation tool in action.

Installation

Prerequisites

• Python 3.8+
• CUDA-capable GPU (recommended for faster processing)
• ffmpeg

Setup Instructions

1. Clone the repository (including SAM2 submodule):

   git clone --recursive https://github.com/yuqunw/scenediff_annotator
   cd scenediff_annotator

2. Create conda environment and install SAM2 and dependencies:

   conda create -n scenediff_annotator python=3.10 -y
   conda activate scenediff_annotator
   pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121 # Install the pytorch fitting your nvcc version    
   pip install -r requirements.txt
   cd sam2 && pip install -e .

3. Download SAM2 checkpoints:

   bash download_ckpts.sh
   cd ../..

For detailed SAM2 installation instructions, refer to the official SAM2 repository.

Usage

Starting the Application

1. Launch the backend server:

   python backend.py

2. Open the web interface: Navigate to http://localhost:5000 in your web browser.

Annotation Workflow

1. Upload Videos: Upload a pair of videos, fill in the Scene Type, and wait for the initialization.

2. Configure Object Attributes:

   • Specify the number of changed objects
   • Specify the name, multiplicity index
   • Specify the appearance in video 1 and video 2 and number of frames for annotation
   • Specify the deformabiltiy

3. Provide Annotations:

   • Select the frame in the video
   • Click to add point prompts

4. Run Offline Propagation:

   • After all annotation, click Start Offline Job to begin mask propagation
   • SAM2 will propagate objects throughout both videos offline. Could close the page.

5. Review & Refine:

   • Navigate to Review Sessions to visualize results
   • Review or refine annotations if needed

Output Format

The uploaded videos and generated outputs are saved at ./uploads and ./results. The annotation tool generates two primary output files:

• inference_data.pkl: Stores initial prompt inputs used for mask propagation
• segments.pkl: Contains the final segmentation masks and metadata

Segments Structure

The segments.pkl file follows this hierarchical structure:

segments = {
    'scenetype': str,                    # Type of scene change
    'video1_objects': {
        'object_id': {
            'frame_id': RLE_Mask         # Run-length encoded mask
        }
    },
    'video2_objects': {
        'object_id': {
            'frame_id': RLE_Mask         # Run-length encoded mask
        }
    },
    'objects': {
        'object_1': {
            'label': str,                # Object label/name
            'in_video1': bool,           # Present in video 1
            'in_video2': bool,           # Present in video 2
            'deformability': str         # 'rigid' or 'deformable'
        }
    }
}

Loading Masks

To convert RLE masks back to tensors:

import torch
from pycocotools import mask as mask_utils

# Load and decode RLE mask
tensor_mask = torch.tensor(mask_utils.decode(rle_mask))

Acknowledgements

This project is built upon the excellent SAM2 repository (Segment Anything Model 2). We gratefully acknowledge their contributions to the computer vision community.

License

See LICENSE for more information.