VPoS-Bench: Video Pointing and Segmentation Benchmark

VPoS-Bench is a challenging out-of-distribution benchmark designed to evaluate the spatio-temporal pointing and reasoning capabilities of video-language models. It covers a diverse set of five real-world application domains, with fine-grained point-level and segmentation annotations that enable robust evaluation of multimodal models under realistic, temporally complex scenarios.

Webpage: VideoMolmo

Paper: VideoMolmo: Spatio-Temporal Grounding meets Pointing

Model: VideoMolmo on Hugging Face

Code: VideoMolmo on Github

---

🌍 Benchmark Overview

VPoS-Bench tests the generalization of models in five diverse real-world scenarios:

1. Cell Tracking
   Track the trajectory of biological entities (e.g., nuclei or cells) across microscopy video frames.

   Applications: developmental biology, disease modeling

2. Egocentric Vision
   Identify and follow objects or hands in first-person camera footage.

   Applications: activity recognition, assistive tech

3. Autonomous Driving
   Point to traffic participants (pedestrians, vehicles, lights) under varying conditions.

   Applications: self-driving systems, urban scene understanding

4. Video-GUI Interaction
   Follow on-screen elements (e.g., cursors, buttons) across software interface recordings.

   Applications: AI-assisted UI navigation, screen agents

5. Robotics
   Track manipulable objects or robotic end-effectors as they interact in structured environments.

   Applications: robot learning, manipulation planning

---

📁 Dataset Structure

The dataset is organized by domain. Each domain folder contains three subdirectories:

• frames/ – Extracted video frames.
• masks/ – Segmentation masks corresponding to frames.
• annotations/ – JSON files containing text descriptions and point-level annotations.

vpos-bench/
├── cell-tracking/
│   ├── frames/             # Extracted video frames (e.g., frame_0001.jpg, ...)
│   ├── masks/              # Segmentation masks per frame (optional)
│   └── annotations/        # Point coordinates + caption in JSON format
│
├── autonomous-driving/
...
---
├──

📁 Annotation Format

Each annotation is keyed by a unique video ID and consists of:

{
  "video_id": {
    "caption": "natural language instruction here",
    "frames": [
      {
        "frame_path": "domain/frames/video_id/frame_00001.jpg",
        "mask_path": "domain/masks/video_id/0.png",
        "points": [[x, y], ...]
      },
      {
        "frame_path": "domain/frames/video_id/frame_00002.jpg",
        "mask_path": "domain/masks/video_id/1.png",
        "points": [[x, y], ...]
      }
    ]
  }
}