import json import os from PIL import Image import torch from torch.utils.data import Dataset from torchvision.ops import box_convert class M2CAI16ToolLocations(Dataset): """Dataset for m2cai16-tool-locations bounding box annotations."""
boxes = target['boxes'].int() labels = target['labels'] class_names = dataset.CLASSES
def __len__(self): return len(self.samples) m2cai16-tool-locations
This dataset is designed for (bounding boxes) in laparoscopic cholecystectomy videos. It contains annotations for 16 tools, including their positions in video frames. 1. Dataset Overview & Utility Purpose : Train object detection models (e.g., YOLO, Faster R-CNN, DETR) to locate surgical instruments in real-time.
# 16 tool classes (example; adjust to your annotation file) CLASSES = [ 'background', 'grasper', 'scissors', 'hook', 'clipper', 'irrigator', 'specimen_bag', 'bipolar', 'hook_electrode', 'trocars', 'stapler', 'suction', 'clip_applier', 'vessel_sealer', 'ligasure', 'ultrasonic', 'other' ] import json import os from PIL import Image
m2cai16-tool-locations/ annotations/ video01.json # or .xml / .txt video02.json frames/ video01/ frame_000001.jpg ... Here’s a robust parser using and torchvision :
path: ./m2cai16-tool-locations train: images/train val: images/val nc: 16 names: ['grasper','scissors','hook','clipper','irrigator','specimen_bag','bipolar','hook_electrode','trocars','stapler','suction','clip_applier','vessel_sealer','ligasure','ultrasonic','other'] This guide gives you a production‑ready starting point for loading, visualizing, converting, and training on the dataset. Adjust class names and annotation JSON structure based on your exact dataset version. Dataset Overview & Utility Purpose : Train object
def __getitem__(self, idx): img_path, ann = self.samples[idx] image = Image.open(img_path).convert('RGB') # Parse annotations: list of [x1, y1, x2, y2, class_id] boxes = [] labels = [] for obj in ann.get('objects', []): x1, y1, x2, y2 = obj['bbox'] # absolute pixel coords label = self.CLASSES.index(obj['class_name']) boxes.append([x1, y1, x2, y2]) labels.append(label) boxes = torch.as_tensor(boxes, dtype=torch.float32) labels = torch.as_tensor(labels, dtype=torch.int64) image_id = torch.tensor([idx]) area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0]) iscrowd = torch.zeros((len(boxes),), dtype=torch.int64) target = { 'boxes': boxes, 'labels': labels, 'image_id': image_id, 'area': area, 'iscrowd': iscrowd } if self.transform: image, target = self.transform(image, target) return image, target Use matplotlib and torchvision.utils.draw_bounding_boxes :
def __init__(self, root_dir, transform=None): self.root_dir = root_dir self.transform = transform self.samples = [] # Collect all (frame_path, annotation_path) pairs ann_dir = os.path.join(root_dir, 'annotations') for ann_file in os.listdir(ann_dir): if not ann_file.endswith('.json'): continue ann_path = os.path.join(ann_dir, ann_file) video_id = ann_file.replace('.json', '') frame_dir = os.path.join(root_dir, 'frames', video_id) with open(ann_path, 'r') as f: annotations = json.load(f) for frame_name, boxes_info in annotations.items(): frame_path = os.path.join(frame_dir, frame_name) if os.path.exists(frame_path): self.samples.append((frame_path, boxes_info))