katha-ai-iiith/vidsitu-grounding

VidSitu Grounding Dataset

Grounding annotations for the VidSitu benchmark, providing bounding box localizations for semantic role entities across video frames.

Each video segment contains 11 uniformly sampled frames from a 10-second movie clip. For each segment, 5 events (Ev1-Ev5) are annotated with verb-specific semantic roles (e.g., Agent, Patient, Instrument), and each grounded role includes:

• Frame indices where the entity is visible
• Bounding boxes (x1, y1, x2, y2) in pixel coordinates
• Text caption describing the entity

Visualization Example

Below is a visualization of v_UZb2NOHPA2A_seg_0_10 (from The Pursuit of Happyness), Event 1 — "play", showing the grounded entities "kid with curly hair" and "basketball" across frames:

Dataset Structure

Column	Type	Description
vid_seg_int	string	Video segment identifier (e.g., v_VYQoxBs5N2A_seg_120_130)
frames	list[Image]	11 JPEG frames (frame_00 to frame_10)
grounding_annotations	string (JSON)	Full grounding annotations (see format below)

Annotation Format

The grounding_annotations column is a JSON string with this structure:

{
  "Ev1": {
    "Arg0 (player)": {
      "Frames": [0, 2, 3, 4, 5, 6, 7],
      "Bboxes": [[x1, y1, x2, y2], ...],
      "caption": "kid with curly hair"
    },
    "Arg1 (game/music)": {
      "Frames": [3, 4, 5, 6, 7],
      "Bboxes": [[x1, y1, x2, y2], ...],
      "caption": "basketball"
    },
    "Arg2 (instrument used to play game/music)": "_",
    "ArgM (location)": "_",
    "ArgM (manner)": "_"
  },
  "Ev2": { ... },
  ...
  "Ev5": { ... }
}

• Roles with value "_" are not grounded (no bounding box annotation).
• Bboxes are in pixel coordinates [x1, y1, x2, y2] matching the frame resolution.
• Frames lists the 0-indexed frame numbers where the entity is visible.

How to Load and Visualize

from datasets import load_dataset
from PIL import Image, ImageDraw
import json

# Load the dataset
ds = load_dataset("katha-ai-iiith/vidsitu-grounding", split="validation")

# Pick a sample
sample = ds[0]
print(f"Segment: {sample['vid_seg_int']}")

# Parse annotations
annotations = json.loads(sample["grounding_annotations"])

# Visualize Event 1 grounding on frame 5
event = "Ev1"
frame_idx = 5
print(f"\nEvent: {event}")

frame = sample["frames"][frame_idx].copy()
draw = ImageDraw.Draw(frame)

colors = ["red", "blue", "green", "orange", "purple"]
color_idx = 0

for role, data in annotations[event].items():
    if not isinstance(data, dict) or "Bboxes" not in data:
        continue  # skip unannotated roles

    if frame_idx in data["Frames"]:
        bbox_pos = data["Frames"].index(frame_idx)
        x1, y1, x2, y2 = data["Bboxes"][bbox_pos]
        color = colors[color_idx % len(colors)]
        draw.rectangle([x1, y1, x2, y2], outline=color, width=3)
        draw.text((x1, y1 - 15), f"{role}: {data['caption']}", fill=color)
        color_idx += 1

frame.save("grounding_example.png")
print(f"Saved visualization to grounding_example.png")

Iterating Over All Grounded Entities

import json
from datasets import load_dataset

ds = load_dataset("katha-ai-iiith/vidsitu-grounding", split="validation")

for sample in ds:
    annotations = json.loads(sample["grounding_annotations"])

    for event_key in ["Ev1", "Ev2", "Ev3", "Ev4", "Ev5"]:
        event = annotations[event_key]

        for role, data in event.items():
            if not isinstance(data, dict):
                continue
            caption = data["caption"]
            n_frames = len(data["Frames"])
            print(f"{sample['vid_seg_int']} | {event_key} | {role} -> {caption} ({n_frames} frames)")

Full VidSitu Dataset

This repository contains grounding annotations for a subset of the validation set. For the complete VidSitu dataset (all splits, videos, and annotations), please refer to:

• VidSitu on India Data Portal: https://india-data.org/dataset-details/0cdb4ba0-19fb-4401-9b4e-d7266e87f0c2
• VidSitu Project Page: https://vidsitu.org/

Citation

If you use this dataset, please cite the VidSitu paper:

@inproceedings{sadhu2021visual,
  title={Visual Semantic Role Labeling for Video Understanding},
  author={Sadhu, Arka and Gupta, Tanmay and Yatskar, Mark and Nevatia, Ram and Kembhavi, Aniruddha},
  booktitle={CVPR},
  year={2021}
}