fh2019ustc / PolySnake

The official code for “Recurrent Generic Contour-based Instance Segmentation with Progressive Learning”, TCSVT, 2024.
MIT License
67 stars 7 forks source link

用官方给的权重来推理,推理结果十分混乱 #11

Closed wjpstudy closed 4 months ago

wjpstudy commented 6 months ago

cuda环境编译成功之后,我分别尝试了

  1. Baidu Cloud上下载了coco预训练模型,并将其放入 $ROOT/data/model/snake/coco/,运行Demo:python run.py --type demo --cfg_file configs/coco_snake.yaml demo_path demo_images ct_score 0.3,
  2. Baidu Cloud上下载了sbd预训练模型,并将其放入 $ROOT/data/model/snake/sbd/,运行Demo:python run.py --type demo --cfg_file configs/sbd_snake.yaml demo_path demo_images ct_score 0.3, 在 $ROOT/demo_out/ 中结果都很十分混乱,如下: image image image image

然后我尝试着提高ct_score 0.5 ct_score 0.7 甚至是 ct_score 0.9 发现这个参数的修改,对结果没有任何的效果,请问我的问题出现在了哪里?期待您的回复,谢谢!

fh2019ustc commented 6 months ago

看起来像是ct_score 没有生效的问题,但是不应该呀,我提交之前测试过。我最近有点忙,等我月底看看。你先debug看看~祝好

wjpstudy commented 6 months ago

已解决。原因是 PolySnake/lib/networks/snake/ICD.py 中的 valid = detection[0, :, 2] >= 0.05 将ct_score写死了,改为valid = detection[0, :, 2] >= cfg.ct_score 就都正常了。谢谢!