video video 717 717 |
|---|
Laparoscopic Surgical Video Dataset for Robotic Vision
An open surgical video sample for robot perception, visual representation learning, and surgical scene understanding.
This dataset contains one unannotated laparoscopic video: 11 minutes 57 seconds at 720 × 576 resolution and 25 fps. It provides visual source material for developing surgical robotics perception pipelines, from self-supervised learning to instrument detection, tracking, and activity recognition after annotation. The sample is released under Apache License 2.0.
Workflow: download → sample frames or clips → annotate → train → evaluate.
Dataset overview
| Attribute | Specification |
|---|---|
| Content | One laparoscopic surgical video |
| Primary file | videos/laparoscopic_sample_001.mp4 |
| Duration | 717.16 seconds |
| Resolution / frame rate | 720 × 576 / 25 fps |
| Encoding | H.264 in MP4; video only |
| File size | 525,231,261 bytes |
| Annotation status | Unannotated |
| Integrity metadata | manifest.json, including SHA-256 |
| License | Apache-2.0 |
The primary file preserves the original recording without re-encoding. The root-level 2023-04-20_2059_VID0001ab.mp4 is an identical copy of the same recording; use the primary path above for a consistent pipeline.
Applications in surgical robotics
| Application | Learning approach | Data preparation |
|---|---|---|
| Visual representation learning | Adapt a pretrained image or video encoder using masked reconstruction or contrastive objectives. | Quality-controlled frames or short clips. |
| Surgical instrument detection and segmentation | Fine-tune a detector or segmentation model to localize visible instruments. | Reviewed bounding boxes or pixel masks, instrument classes, and visibility labels. |
| Instrument tracking | Associate instrument instances and estimate image-plane motion over time. | Consecutive frames, persistent IDs, and optional tip or shaft keypoints. |
| Surgical activity recognition | Train temporal models on observable actions and transitions. | Timestamped segments with a consistent action vocabulary. |
| Vision-language learning | Learn scene descriptions, retrieval, or question answering from image-text or video-text pairs. | Expert-reviewed captions and questions grounded in visible content. |
The release supplies the raw visual observations. The task-specific labels in this table describe the next preparation step for supervised training.
Quick start
Install the Hugging Face CLI and FFmpeg, then download the primary video and supporting files:
hf download hkust-aaron/laparoscopic-robotic-vision-sample \
videos/laparoscopic_sample_001.mp4 manifest.json LICENSE README.md \
--repo-type dataset --local-dir surgical-video
mkdir -p surgical-video/frames
ffmpeg -i surgical-video/videos/laparoscopic_sample_001.mp4 \
-vf fps=2 -q:v 2 surgical-video/frames/frame_%06d.jpg
This example samples at 2 fps, producing approximately 1,434 images for initial review and annotation. Use native-rate consecutive frames for tracking or faster motion. Maintain a manifest mapping each extracted image to its source frame index and presentation timestamp.
From video to training data
Prepare and annotate. Check decoding, timestamps, blur, occlusion, and exposure. Define instrument classes and rules for partially visible objects before labeling. For detection, store bounding boxes; for segmentation, masks; for tracking, persistent instance IDs. An annotation record can include:
video_id, frame_index, timestamp_seconds, image_path,
class_id, bbox_or_mask, instrument_id, keypoints, visibility
For temporal labels, use start_seconds, end_seconds, and action_label. For vision-language tasks, pair reviewed captions with explicit frame or clip references. Surgical interpretation benefits from expert review.
Train a baseline. Start with a pretrained visual encoder and a task-specific head. Compare a frozen-encoder baseline against partial or full fine-tuning. Preserve aspect ratio with resizing and padding; keep geometric augmentation synchronized with boxes, masks, and keypoints. For temporal experiments, sample clips of 16 or 32 frames and record their sampling rate and stride. Save the model version, preprocessing configuration, label schema, split definition, and random seed for reproducibility.
Evaluate and expand. Separate training and validation into contiguous temporal blocks, leaving a gap to reduce leakage between neighboring frames. Use detection mAP, segmentation Dice/IoU, tracking HOTA/IDF1, or activity-recognition macro-F1 and segment-level F1 as appropriate. Within-video evaluation tests pipeline behavior; additional independent recordings across procedures, patients, devices, and institutions enable cross-case generalization studies. Review errors by visibility, motion blur, and illumination to guide further data collection.
Toward action-conditioned robot learning
The video can support development of the visual observation encoder in a larger robot-learning system. Extending that system to policy learning requires synchronized demonstrations pairing images with robot state and executed actions:
observation_t = {image_t, joint_state_t, end_effector_pose_t, ...}
action_t = {commanded_motion_t, gripper_command_t, ...}
Collect camera calibration, coordinate-frame definitions, timestamp synchronization, action units, and control frequency, together with force or contact signals where relevant. These paired demonstrations support behavior cloning and other imitation-learning experiments, followed by simulation and task-specific validation.
Release notes and provenance
The contributor supplied the recording and confirmed permission for public distribution, commercial AI training, and release under Apache License 2.0. Acquisition and procedure metadata can be added as they are verified. Expanded collections should document provenance, consent, and privacy review alongside technical metadata.
The original MP4 decoded in an initial check, with an FFmpeg wrong sample count container warning. Verify decoded frames and timestamps in the chosen preprocessing pipeline and retain the result in processing logs.
Broader surgical video resources
This repository provides one public sample from a broader initiative. The team reports access to approximately 50,000 hours of surgical video resources and welcomes discussions about data requirements for medical AI and surgical robotics research. Availability, scope, delivery, and licensing terms for those resources are agreed separately; the Apache-2.0 license applies to the sample published in this repository.
Contact
For dataset inquiries, broader video resources, and collaboration, contact aaron@aaronxu.com.
- Downloads last month
- 41