search

lkeab / BCNet

Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers [CVPR 2021]
https://arxiv.org/abs/2103.12340
MIT License
529 stars 74 forks source link

Visualizaiton result is pretty bad. #124

Closed willpat1213 closed 1 year ago

willpat1213 commented 1 year ago

Thanks for such great job! I have some trouble in visualization. Run the code "bash visualization.sh" to visualize segm results.

CUDA_VISIBLE_DEVICES=0 python3 demo/demo.py --config-file configs/fcos/fcos_imprv_R_101_FPN.yaml --input ./demo/test_img.jpg --output ./demo/result_img --opts MODEL.WEIGHTS ./pretrained_models/model_final_beta.pth

The visualize results as: 000000000285

The repo refer to https://github.com/Shadercloud/detectron2 The weight refer to https://hkustconnect-my.sharepoint.com/:u:/g/personal/lkeab_connect_ust_hk/EfiDFLLEawFJpruwuOl3h3ABBjAKysTf0qJQU80iaKbqYg?e=igzC51 I have no idea if need to train the model then visualize.

lkeab commented 1 year ago

because you use the repo provided by The repo refer to https://github.com/Shadercloud/detectron2, you should also employ the pretrained model by it. Our provided model is using FCOS detector.

willpat1213 commented 1 year ago

I took a look at the config and code of this repo, which also uses fcos, but it seems that it does not match the weight model you provided. Is it because the version of D2 is different or the implementation is different?

The details of the visualization process are shown in the figure:

image

The infer.sh: 

CUDA_VISIBLE_DEVICES=0 python3 demo/demo.py --config-file configs/fcos/fcos_imprv_R_101_FPN.yaml \
  --input ./demo/test_img/ \
  --output ./demo/result_img/ \
  --opts MODEL.WEIGHTS ./pretrained_models/model_final_beta.pth