← Computer Vision · Repository
The detection projects implement detector-specific TensorFlow models and local train/test tooling. They share a practical pattern: dataset adapters, configuration-driven roots, model/loss code, training loops, inspection, and bounded local caches. COCO, Pascal VOC, and ImageNet-related paths appear in the source; availability is a local prerequisite, not a bundled dataset.
| Detector family | Implementation | Notes |
|---|---|---|
| R-CNN | 1_RCNN |
Region-based detection pipeline |
| Fast R-CNN | 2_FastRCNN |
ROI classification/regression path |
| Faster R-CNN | 3_FasterRCNN |
RPN, proposal, ROI and detector training code |
| Mask R-CNN | 4_MaskRCNN |
Detection/mask-oriented extension |
| YOLO | 5_YOLOv1, 7_YOLOv2, 8_YOLOv3 |
Grid/anchor-style one-stage detectors |
| SSD | 6_SSD |
Multi-scale one-stage detector |
Several detectors use tf.data, custom tf.keras.layers.Layer/models,
tf.GradientTape, and device-aware distribution setup. Some folders include
test entry points; Fast R-CNN also contains tests/test_run_controls.py.
Generated TensorBoard reports, visual inspections, checkpoints, and caches are
run artifacts—not shipped benchmark results.