Zero-shot demo

[GioFic95]

I propose a zero-shot demo that can be fully executed via command line.

You can run it with:

python demo_zero.py --img_path <input file or dir path> --out_path <output_path> --zero_shot --two_passes --skip_train --model_name DAVE_0_shot --model_path material --backbone resnet50 --swav_backbone --reduction 8 --num_enc_layers 3 --num_dec_layers 3 --kernel_dim 3 --emb_dim 256 --num_objects 3 --num_workers 8 --use_objectness --use_appearance --batch_size 1 --pre_norm

where:

• --img_path is the path of the image or directory of images to use as input;
• --out_path is the output path where the result(s) will be saved;
• --two_passes enables the 2-passes strategy as suggested by @jerpelhan in https://github.com/jerpelhan/DAVE/issues/5#issuecomment-2197006231.

If you prefer to show the result instead of saving it, just run:

python demo_zero.py --img_path <input file or dir path> --show --zero_shot --two_passes --skip_train --model_name DAVE_0_shot --model_path material --backbone resnet50 --swav_backbone --reduction 8 --num_enc_layers 3 --num_dec_layers 3 --kernel_dim 3 --emb_dim 256 --num_objects 3 --num_workers 8 --use_objectness --use_appearance --batch_size 1 --pre_norm

with --show instead of --out_path.

[jerpelhan]

Great! Thank you so much for contributing to this repository. Did you try setting bounding boxes to torch.zeros((1,3,4)), as only the shape of bounding boxes sent to this function is important? This way no modification is needed in the dave.py.

[GioFic95]

Thank you for the hint, I applied what you suggested in the new commit dc5f2d3 (#9).