Object Detection with PE
Getting started
Please refer to INSTALL.md for installation and dataset preparation instructions.
Results and Fine-tuned Models
LVIS
| detector | vision encoder | box AP |
mask AP |
download |
|---|---|---|---|---|
| Mask R-CNN | PE core G | 51.9 | 47.9 | model |
| Mask R-CNN | PE spatial G | 54.2 | 49.3 | model |
COCO
| detector | vision encoder | box AP |
mask AP |
download |
|---|---|---|---|---|
| Mask R-CNN | PE core G | 57.0 | 49.8 | model |
| Mask R-CNN | PE spatial G | 57.8 | 50.3 | model |
Training
By default, we use 64 GPUs in slurm training, for example
sbatch scripts/coco/train_mask_rcnn_PEspatial_G_coco36ep.sh
Evaluation
Evaluation is running locally
bash scripts/evaluate_local.sh --config-file projects/ViTDet/configs/COCO/mask_rcnn_PEspatial_G_coco36ep.py train.output_dir="/path/to/output_dir" train.init_checkpoint="/path/to/mask_rcnn_PEspatial_G_coco36ep.pth"
SOTA COCO Object Detection
| detector | vision encoder | box AP |
box(TTA) AP |
download |
|---|---|---|---|---|
| DETA | PE spatial G | 65.2 | 66.0 | model |
More details are in DETA_pe
Acknowledgment
This code is built using detectron2 and DETA.