EVALUATION.md

Evaluation - Few-shot Instance Segmentation

Prepare data

Download following datasets:

COCO 2014

• Download cocosplit.

VOC

• Download voc-cocostyle.zip.

Create a directory 'datasets' for the above datasets in 'inference_fsod/' and appropriately place each dataset to have following directory structure:

datasets/
├── coco/           
│   ├── annotations/
│   │   └── instances_{train,val}2014.json
│   └── {train,val}2014/
├── cocosplit/
│   ├── datasplit/
│   ├── seed0/
│   ├── seed1/
│   └── ...
├── VOCOutput/
│   ├── annotations
│   │   ├── train.json
│   │   ├── val.json
│   │   └── val_converted.json
│   ├── train/
│   ├── val/

Testing

cd inference_fsod

python tools/preprocess.py \
  --config-file configs/COCO/1shots/seed0.yaml \
  --opts OUTPUT_DIR /path/to/fsod/coco/1shot/seed0 \
  MODEL.SINE.preprocess True \
  MODEL.WEIGHTS /path/to/pt_sine/pytorch_model.bin

python tools/test.py \
  --num-gpus=8 \
  --config-file configs/COCO/1shots/seed0.yaml \
  --opts MODEL.SINE.preprocess False \
  OUTPUT_DIR /path/to/fsod/coco/1shot/seed0 \
  MODEL.WEIGHTS /path/to/pt_sine/pytorch_model.bin