Small Object Detection HPO
Environment: torch.dev.gpu Status: Planned -- datasets in lakehouse Datasets: landing/coco/, landing/visdrone/, landing/bdd/
Overview
Training and fine-tuning object detection models (YOLO, RT-DETR, D-FINE, DETR) on the Auraison platform using datasets already consolidated in the lakehouse landing tier.
Datasets
| Dataset | Size | Objects | Description |
|---|---|---|---|
| COCO 2017 | ~49 GiB | detection, captions, segmentation, keypoints | MS COCO benchmark |
| VisDrone 2019 | ~96 GiB | DET, MOT, SOT, VID | Drone-view multi-task benchmark |
| BDD 100K | ~652 GiB | 100k driving images (train/val/test) | Berkeley driving dataset |
Rerun visualizations
VisDrone video sequences with bounding box overlays are viewable in the Rerun web viewer via Cloudflare R2 public artifacts.
Platform services
- Data plane: Datasets in landing/, training outputs + model checkpoints in warehouse/, Rerun recordings in R2
- User plane: Training jobs as Ray Jobs on torch.dev.gpu
- Control plane: NotebookAgent for job submission, WandBAgent for experiment tracking