Official code of the paper "Expanding Low-Density Latent Regions for Open-Set Object Detection" (CVPR 2022)

@csuhan 您好，我想请教论文中的一个点。 Figure 2. t-SNE visualization of latent features. 这里提到彩色的点是VOC类别（已知），而黑色三角点则是非VOC类别（未知，取自COCO）

在您的工作中，您将未知类别设定为数量为1的1个类别（而不是更多数量），这样训出来的模型，我们就很自然地认为，未知类别也会聚成一个簇，就像图2 (b)中一样。同时还有一些离散点，分散在各个已知类别簇中。

但是实际上来看，1个未知类别，它应该蕴含了众多潜在的类别，例如COCO类别数量-VOC类别数量=80-20=60，也就是1个未知类别可能就蕴含了潜在的60个类别。而将60个类别的特征聚集在了一起，形成了图2 (b)，这是不是有点奇怪？也就是想问，只有1个未知类的类中心，是不是不太合理？

想请教您的看法，谢谢！

Nice job! I am trying to reproduce your work. But I find that loss_cls_ic is 0 for most of the time after the training started. Is it normal? (I set batch_size=4 because of the limited computational resources.) Thanks.

When I followed the README to install opendet2, I got some trouble. Here is my command. I have several rtx3090 gpus.

# CUDA V11.1
# torch 1.9.0
# python 3.8
conda create -n opendet2 python=3.8 -y
conda activate opendet2
# get pytorch 1.9.0. I got RuntimeError: CUDA error: device-side assert triggered when using torch 1.8.1
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
# build opencv
pip install opencv-python
pip install opencv-contrib-python
# build detectron2. DO NOT build detectron2 from latest SOURCE. In the latest version, some named methods have been removed.
pip install detectron2==0.5 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.9/index.html
# build opendet2
cd opendet2
pip install -v -e .

When I run the last command pip install -v -e ., I got these error message:

(opendet2) yupeng@ubuntu:~/opendet/opendet2$ pip install -v -e .
Using pip 21.2.4 from /home/yupeng/anaconda3/envs/opendet2/lib/python3.8/site-packages/pip (python 3.8)
Looking in indexes: https://mirrors.bfsu.edu.cn/pypi/web/simple/
Obtaining file:///home/yupeng/opendet/opendet2
    Running command python setup.py egg_info
    error: Multiple top-level packages discovered in a flat-layout: ['demo', 'configs', 'opendet2', 'datasets', 'detectron2'].

    To avoid accidental inclusion of unwanted files or directories,
    setuptools will not proceed with this build.

    If you are trying to create a single distribution with multiple packages
    on purpose, you should not rely on automatic discovery.
    Instead, consider the following options:

    1. set up custom discovery (`find` directive with `include` or `exclude`)
    2. use a `src-layout`
    3. explicitly set `py_modules` or `packages` with a list of names

    To find more information, look for "package discovery" on setuptools docs.
WARNING: Discarding file:///home/yupeng/opendet/opendet2. Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.
ERROR: Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.
(opendet2) yupeng@ubuntu:~/opendet/opendet2$

I could use pip install setuptools==58.2.0 and retry pip install -v -e ., then everything is fine. It seems there are some problems by using the lateset setuptools>=61.0. Maybe you can find more information in the link below: https://github.com/pypa/setuptools/issues/3197 https://github.com/pypa/setuptools/issues/3227 https://github.com/facebookresearch/detectron2/issues/3943 https://github.com/facebookresearch/detectron2/issues/3811 Good luck.

使用基本的resnet backbone的faster rcnn会发生死锁。我简单的把Base_RCNN_FPN.yaml换成了detectron2中的Base_RCNN_C4.yaml。 使用readme中示例代码训练时卡在训练第一个batch的地方，GPU占用率100%，但是显存只占了2400M，一夜过去14小时还是卡在该位置，没有任何输出或报错。改为单GPU训练正常，可以提供一些帮助吗？

When I run train_net.py, I get this issue after loading R-50.pkl. I want to know how to solve that. Thanks a lot. My environment: CUDA 11.1, python 3.7, torch 1.8.1

opendet2/solver/build.py line 39: 注释和代码不符，注释中bias的weight_decay为默认值，实际代码中被设为None，导致以下错误： TypeError: add(): argument 'alpha' must be Number, not NoneType

Hi,

Amazing work on open-set detection! I trained the model after doing the dataset separation steps you suggest, and with exact same configs. The only difference is that I used 1 GPU instead of 8 GPUs, and these are the results I obtained. Interestingly, WI and AOSE metrics are worse, but AP is better. Do you think this much difference is expected just from using fewer GPUs, or is there some other issue I need to look for? Thanks in advance.

VOC-COCO-20 Result | WI ↓ | AOSE ↓ | AP u↑ -- | -- | -- | -- Paper | 14.95 | 11286 | 14.93 Reproduced | 20.68 | 13370 | 21.36

VOC-COCO-0.5n Result | WI ↓ | AOSE ↓ | AP u↑ -- | -- | -- | -- Paper | 6.44 | 3944 | 9.05 Reproduced | 55 | 5369 | 18.09

Running command python setup.py egg_info error: Multiple top-level packages discovered in a flat-layout: ['data', 'engine', 'solver', 'config', 'modeling', 'evaluation'].