Reproduce Instructions
April 12, 2024 ยท View on GitHub
main results
To reproduce the main results:
OV-COCO
## Detic on the OV-COCO with a ResNet-50 backbone
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml
## CoDet on the OV-COCO with a ResNet-50 backbone
python train_net.py --eval-only --config-file configs/CoDet_RN50_COCO.yaml
OV-LVIS
## Detic on the OV-LVIS with a SwinB backbone
python train_net.py --eval-only --config-file configs/Detic_SwinB_LVIS.yaml
## CoDet on the OV-LVIS with a ResNet-50 backbone
python train_net.py --eval-only --config-file configs/CoDet_RN50_LVIS.yaml
## Codet on the OV-LVIS with a SwinB backbone
python train_net.py --eval-only --config-file configs/CoDet_SwinB_LVIS.yaml
## Codet on the OV-LVIS with an EVA02 backbone
python train_net.py --eval-only --config-file configs/CoDet_EVA02_LVIS.yaml
ablation studies
To reproduce the ablation studies, you can set the OVERLAP_TOPK, ALPHA and BETA to 0 to disable the ARP LQ, AOC VS and AOC LQ respectively.
## baseline
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=0 MODEL.ALPHA 0.0 MODEL.BETA 0.0
## ARP LQ (Aggregated Region-Proposal -- Localization Quality Estimation)
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=3 MODEL.ALPHA 0.0 MODEL.BETA 0.0
## AOC VS (Aggregated Object-Classification -- Visual Similarity Estimation)
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=0 MODEL.ALPHA 0.05 MODEL.BETA 0.0
## AOC LQ (Aggregated Object-Classification -- Localization Quality Estimation)
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=0 MODEL.ALPHA 0.0 MODEL.BETA 0.75
## ARP LQ + AOC VS
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=3 MODEL.ALPHA 0.05 MODEL.BETA 0.0
## ARP LQ + AOC LQ
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=3 MODEL.ALPHA 0.0 MODEL.BETA 0.75
### AOC VS + AOC LQ
python train_net.py --eval-only --config-file configs/Detic_RN50_COCO.yaml \
MODEL.OVERLAP_TOPK=0 MODEL.ALPHA 0.05 MODEL.BETA 0.75
core functions
The core functions of ARP LQ, AOC VS and AOC LQare as follows:
ARP LQfor Faster R-CNN: customrpn.py L235-L240ARP LQfor CenterNet2: centernet.py L659-L675AOC VS(unified) : zero_shot_classifier.py L121-L125AOC LQfor Faster R-CNN: aggdet_fast_rcnn.py L111-L117AOC LQfor CenterNet2: aggdet_roi_heads.py L162-L164