Kavitha/kathakali-test

Kathakali Test

A small video dataset of Kathakali dance recordings with time-stamped segment annotations, intended for testing data pipelines for segment classification.

Note: the segment labels in this version are dummy data (random animal names) used only to test the dataset structure. They do not describe the content of the videos.

Dataset summary

Videos	5
Total duration	~30 minutes
Frame rate	60 fps
Segments	212 (172 train / 40 test)
Classes	10

Dataset structure

├── README.md
├── videos_hd/
│   ├── metadata.jsonl
│   └── IMG_0781.mp4 ...
└── videos_sd/
    ├── metadata.jsonl
    ├── segments_train.csv
    ├── segments_test.csv
    └── IMG_0781.mp4 ...

Configurations

Config	Content	Splits
videos (default)	SD videos, 1024 px high, one row per video	train (all videos)
videos_hd	HD videos at original resolution (3840×2160), one row per video — not yet uploaded	train (all videos)
segments	Labelled time segments, one row per segment, each with its SD video	train, test

The SD and HD folders contain the same videos with the same file names, so segment annotations apply to both.

The segment CSVs sit next to the videos because the Hugging Face video loader only links a CSV's file_name column to videos in the same folder.

Videos

All videos are H.264 MP4 with no audio track. They were recorded on an iPhone in HDR (HLG, BT.2020) and converted to standard dynamic range (BT.709).

metadata.jsonl fields:

Field	Type	Description
file_name	string	Video file name
video_id	string	File name without extension; key used by the annotations
duration_sec	float	Duration in seconds
fps	float	Frame rate
num_frames	int	Number of frames
width, height	int	Resolution in pixels

Segments

Field	Type	Description
video	video	The full SD video the segment comes from (stored as file_name in the CSV)
video_id	string	Video the segment belongs to
start_time, end_time	string	Segment boundaries as HH:MM:SS.mmm
start_sec, end_sec	float	Segment boundaries in seconds
start_frame, end_frame	int	Segment boundaries as frame indices (0-based, seconds × fps rounded)
label	string	Class label

Example row:

file_name,video_id,start_time,end_time,start_sec,end_sec,start_frame,end_frame,label
IMG_0781.mp4,IMG_0781,00:00:02.099,00:00:04.174,2.099,4.174,126,250,Peacock

Labels (10): Deer, Elephant, Horse, Lion, Monkey, Parrot, Peacock, Snake, Swan, Tiger.

Splits: segments are split per label, with 20% of each label's segments in test, so both splits have the same class balance (17–18 train and 4 test segments per label). Train and test segments can come from the same video.

Usage

Requires datasets 4.0 or later and torchcodec for video decoding.

from datasets import load_dataset

segments = load_dataset("Kavitha/kathakali-test", "segments", split="train")

# each segment row comes with its video; decode just the segment's frames
seg = segments[0]
clip = seg["video"].get_frames_played_in_range(seg["start_sec"], seg["end_sec"])

print(seg["label"], clip.data.shape)  # e.g. Peacock torch.Size([126, 3, 1024, 1820])

Videos are downloaded once and shared by all segments from the same video; frames are only decoded for the requested range.

Once the HD videos are uploaded, load them for full-resolution frames and join them to the segments by video_id:

videos_hd = load_dataset("Kavitha/kathakali-test", "videos_hd", split="train")
idx = {vid: i for i, vid in enumerate(videos_hd["video_id"])}
clip = videos_hd[idx[seg["video_id"]]]["video"].get_frames_played_in_range(seg["start_sec"], seg["end_sec"])

Dataset creation

Source data

Processing

1. Raw iPhone .MOV recordings (HEVC, 10-bit HDR) were converted to H.264 MP4 with ffmpeg, tone-mapped from HLG to BT.709, and the audio removed.
   • SD: scaled to 1024 px height, keeping the aspect ratio.
   • HD: original resolution.
2. Annotations were made as .srt subtitle files (one per video, named after the video) and converted to the segment CSVs.

Annotations

Personal and sensitive information

The videos show identifiable people performing.

Limitations

• Segment labels are currently placeholders, not real annotations.
• Small dataset (5 videos), so not suitable for training models that need to generalise.
• Train and test segments share videos, so test scores may be higher than on unseen recordings.

License

Citation