search

HRNet / HRNet-Object-Detection

Object detection with multi-level representations generated from deep high-resolution representation learning (HRNetV2h). This is an official implementation for our TPAMI paper "Deep High-Resolution Representation Learning for Visual Recognition". https://arxiv.org/abs/1908.07919
Apache License 2.0
643 stars 98 forks source link

w40 model's map=0.000 #22

Closed zkunob closed 4 years ago

zkunob commented 5 years ago

the model doesn't work,all the ap values are 0.000

wondervictor commented 5 years ago

Please provide more details.

zkunob commented 5 years ago

I test the model on coco with the method of mmdetection，but all the test value is 0

zekun0227 邮箱：zekun0227@163.com

Signature is customized by Netease Mail Master

On 09/08/2019 09:59, Tianheng Cheng wrote:

Please provide more details.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

wondervictor commented 4 years ago

May you provide your 'config' file. Our code has some modifications compared with the official mmdetection.

DangerousY commented 4 years ago

May you provide your 'config' file. Our code has some modifications compared with the official mmdetection.

I meet with the same problem, i have 10 classes and i use faster_rcnn_hrnetv2p_w18_1x.py. I change it a little. This is my config file

5e1fbe4b3eaefca7fe8b1d87f3b3938 9b54d7473c5917173ca006b89cbc7f1 399d23ed0c571ce35fa867fae08fbf1 dd17adc51f412042fcd6f3fce00d29f

wondervictor commented 4 years ago
  1. You didn't change the num_classes in config file, it still outputs 81 classes
  2. Pretrained model is necessary. We didn't test training w/o pre-trained model because the FROZEN BN is not correctly initialized.
  3. If those mentioned above cannot solve your problem, please check your data and your annotations. We can not guarantee that your data and the corresponding annotations are absolutely reliable.

Confidently, we have tested out code in both training and testing phases and we can obtain better results than ResNets and ResNeXts.

DangerousY commented 4 years ago
  1. You didn't change the num_classes in config file, it still outputs 81 classes
  2. Pretrained model is necessary. We didn't test training w/o pre-trained model because the FROZEN BN is not correctly initialized.
  3. If those mentioned above cannot solve your problem, please check your data and your annotations. We can not guarantee that your data and the corresponding annotations are absolutely reliable.

Confidently, we have tested out code in both training and testing phases and we can obtain better results than ResNets and ResNeXts.

Thanks, it works!

wondervictor commented 4 years ago

Solved, close