GFocalV2
GFocalV2 copied to clipboard
implus
Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection, CVPR2021
Hello. How can I install this repo and run training? Clicking on INSTALL.md and GETTING_STARTED.md send on page with 404 error.
你好，感谢开源这么有意义的工作，我这边有点关于DCN的具体配置的疑问，如网页上所述： _[3] dcnv2 denotes deformable convolutional networks v2. Note that for ResNe(X)t based models, we apply deformable convolutions from stage c3 to c5 in backbones._ 训练时使用的都是DCNv2，并在c3-c5都替换为DCNv2。但我看config/gfocal/目录下的脚本都是如下设置： _dcn=dict(type='DCN', deform_groups=1, fallback_on_stride=False), stage_with_dcn=(False,...
(open-mmlab) fsr@titanxp:~/FoveaBox-01$ python tools/train.py configs/foveabox/fovea_r50_fpn_gfl.py Traceback (most recent call last): File "tools/train.py", line 9, in from mmdet.apis import (get_root_logger, init_dist, set_random_seed, File "/home/fsr/FoveaBox-01/mmdet/apis/__init__.py", line 2, in from .inference import (inference_detector,...
 Wonderful work! I have a question about the computation of quality focal loss. I notice the negative part of the quality focal loss is computed via the whole indices...
Thanks for the nice work. I checked GFLv2 paper, this repo and config files, but I could not find which scales you used to obtain 53.3 AP on R2-101-DCN backbone....
翔哥你好，我重新了试了以下gflv2，的确提点还是很多的，并且我将topk4操作换成了普通卷积，也没有发现明显的掉点，现在已经适合部署了。 但是我发现如果我把默认的GN换成BN，fintune的时候loss和gn没有太大差别，但是map就只有0.0几了，请问这是什么情况？ 我把shared head 改成 private head，发现map又正常了。请问到底是啥情况
训练集中包含了一些负样本（没有检测目标）的图像， 为了训练时加入这些图像，我设置了 filter_empty_gt=False,如下： ` train=dict( type=dataset_type, ann_file=data_root + '/annotations/LumptrainCOCOmmdet.json', img_prefix=data_root+'/train', classes=classes, filter_empty_gt=False, pipeline=train_pipeline),` 训练了几步后，报了以下错误： `Consider using one of the following signatures instead: nonzero(*, bool as_tuple) (Triggered internally at /opt/conda/conda-bld/pytorch_1595629427478/work/torch/csrc/utils/python_arg_parser.cpp:766.) &...
你好，我想基于centernet这种网络加上GFlcalV2的loss来训练，但是发现centernet只有wh这个可以类比于gfocalhead这个部分，是不是distrib 这个分支可以改成2n的通道？ 另外，如果是360度旋转的目标检测任务，GFocalV2还能wok吗？
