joonhaim/surgical-tool-recognition-full-multiview

---
pretty_name: Surgical Tool Recognition Full Multiview
task_categories:
- object-detection
language:
- en
license: mit
tags:
- medical
- computer-vision
- surgery
- object-detection
- yolo
size_categories:
- 1K<n<10K
---

# Surgical Tool Recognition Full Multiview

## Summary

This dataset contains images of individual surgical instruments for object detection.
It was originally created in YOLO format and exported here to a Hugging Face-friendly structure with `metadata.jsonl` files for each split.

## Splits

- train: 2016
- validation: 252
- test: 252

Total: 2520 images

## Classes

- `0` = clamp
- `1` = needle_holder
- `2` = scalpel
- `3` = shear
- `4` = tweezer

## File naming convention

Each image filename follows:

`{TOOL}_{VIEW}_{BACKGROUND}_{STATE}_{index}`

Example:

`CM_CLO_BL_CL_003.jpg`

### Tool codes
- `CM` = clamp
- `NH` = needle holder
- `SC` = scalpel
- `SH` = shear
- `TW` = tweezer

### View codes
- `CLO` = close view
- `OBL` = oblique view
- `TOP` = top view

### Background codes
- `BL` = blue
- `WH` = white
- `TR` = metal tray
- `GR` = green

### State codes
- `OP` = open
- `CL` = closed
- `NA` = not applicable

Hinged instruments (`CM`, `NH`, `SH`) use `OP` and `CL`.
Non-hinged instruments (`SC`, `TW`) use `NA`.

## Structure

- `train/images/` + `train/metadata.jsonl`
- `validation/images/` + `validation/metadata.jsonl`
- `test/images/` + `test/metadata.jsonl`

## Annotation format

The original annotations were in YOLO format:

`class_id x_center y_center width height`

In this export, annotations are stored in `metadata.jsonl` with:
- `file_name`
- `width`
- `height`
- `stem`
- `split`
- `objects.bbox`
- `objects.categories`
- `objects.category_names`

Bounding boxes are stored as:

`[x_min, y_min, width, height]`

## Notes

- One instrument per image
- One bounding box per image
- Controlled viewpoints and backgrounds
- Intended for research and educational use in surgical computer vision