aselimc Detection & Segmentation Pipeline
Data Formats
- COCO: JSON annotations,
[x, y, w, h]bbox format - VOC: XML per image,
[xmin, ymin, xmax, ymax] - YOLO: txt per image,
[class cx cy w h]normalized
Model Selection
| Task | Recommended | Alternative |
|---|---|---|
| 2D Detection | RT-DETR, YOLOv8 | Faster R-CNN, DINO |
| Instance Seg | Mask2Former | Mask R-CNN |
| Semantic Seg | SegFormer, Mask2Former | DeepLabv3+ |
| Panoptic | Mask2Former | - |
| 3D Detection | CenterPoint, PointPillars | VoxelNet |
Evaluation
- Detection: COCO mAP (AP@0.5, AP@0.5:0.95, AP_S/M/L)
- Segmentation: mIoU, per-class IoU, boundary F1
- Always report FLOPs and inference latency alongside accuracy
Key Libraries
mmdetection, detectron2, ultralytics, torchvision