Cascade-DETR: Delving into High-Quality Universal Object Detection
Mingqiao Ye, Lei Ke, Siyuan Li, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, Fisher Yu
ETH Zurich & HKUST
We organize the folder as follows.
cascade_dn_detr
|____data
|____datasets
|____models
|____util
|____work_dir
|____engine.py
|____main.py
We organize all datasets in the data folder. You can check datasets/coco.py for corresponding json annotation names.
data
|____coco
|____uvo
|____cityscapes
|____brain_tumor
|____document_parts
|____smoke
|____egohands
|____plantdoc
|____bdd100k
|____people_in_paintings
We provide a script train.sh for training with multi gpus. Note that required packages and deformable attention need to be installed. For small datasets, you can modify --epochs to 50 and --lr_drop to 40.
bash train.sh [dataset_file] [num_gpus]
For example
bash train.sh coco 8
We provide a script test.sh for training with multi gpus.
bash test.sh [dataset_file] [num_gpus] [checkpoint_path]
For example
bash test.sh coco 8 pretrained_checkpoints/coco.pth
We provide pre-trained checkpoints with ResNet50 backbone on UDB10 benchmark. As described in our paper, we use 12-epoch setting for large datasets and 50-epoch setting for small datasets.
| Dataset | Epochs | AP | Checkpoint |
|---|---|---|---|
| COCO | 12 | 45.5 | Link |
| UVO | 12 | 28.4 | Link |
| Cityscapes | 50 | 29.0 | Link |
| Brain tumor | 50 | 46.5 | Link |
| Document Parts | 50 | 50.9 | Link |
| Smoke | 50 | 71.8 | Link |
| Egohands | 12 | 77.6 | Link |
| PlantDoc | 50 | 49.1 | Link |
| BDD100K | 12 | 30.2 | Link |
| People in Painting | 50 | 13.4 | Link |