Download the model weights of DINOv2, SAM and Semantic-SAM, and organize them as follows.
models/
dinov2_vitl14_pretrain.pth
sam_vit_h_4b8939.pth
swint_only_sam_many2many.pth
You can test one-shot semantic segmentation performance of Matcher on COCO-20i, run:
python main_oss.py \
--benchmark coco \
--nshot 1 \
--max_sample_iterations 64 \
--box_nms_thresh 0.65 \
--sample-range "(1,6)" \
--topk_scores_threshold 0.0 \
--use_dense_mask 1 \
--use_points_or_centers \
--purity_filter 0.02 \
--iou_filter 0.85 \
--multimask_output 1 \
--sel_stability_score_thresh 0.90 \
--use_score_filter \
--alpha 1.0 --beta 0. --exp 0. \
--num_merging_mask 9 \
--fold 0 --log-root "output/coco/fold0"
- You can replace
--benchmark coco
with--benchmark lvis
to test LVIS-92i. - You can replace
--nshot 1
with--nshot 5
and replace--num_merging_mask 9
with--num_merging_mask 5
to test 5-shot performance on COCO-20i. - You can find more commands in
scripts/
for other datasets.
Launch the local demo built with gradio:
python app.py