search

NVlabs / AL-MDN

Official pytorch implementation of Active Learning for deep object detection via probabilistic modeling (ICCV 2021)
https://openaccess.thecvf.com/content/ICCV2021/html/Choi_Active_Learning_for_Deep_Object_Detection_via_Probabilistic_Modeling_ICCV_2021_paper.html
Other
167 stars 23 forks source link

how to use gmm in other object detection model? #2

Closed Shanyaodedanshen closed 3 years ago

Shanyaodedanshen commented 3 years ago

Thx for ur work. How to use gmm in other object detection model? Such as RetinaNet, ATSS, GFLv2.

jwchoi384 commented 3 years ago

Hello, I haven't tried our proposed modeling for the algorithm you mentioned. However, I think that the proposed method can be applied to other networks because most of the output of object detection algorithm is the same. It worked well when I applied it to Faster-RCNN.

Shanyaodedanshen commented 3 years ago

Hello, I haven't tried our proposed modeling for the algorithm you mentioned. However, I think that the proposed method can be applied to other networks because most of the output of object detection algorithm is the same. It worked well when I applied it to Faster-RCNN.

  • First, increase the number of output size of both classification and localization heads by considering the number of Gaussian mixture model (GMM) components.
  • Second, modify the loss function as the proposed loss function with the GMM parameters of each head.
  • Lastly, after training is completed, modify the calculation of bounding box coordinates and classification information considering GMM in inference.

Thank you. I have apply your work in RetinaNet. But When I use the raw SSD+GMM for BDD dataset, mAP with only 2.+, which use the same parameters with COCO. And when I use RetinaNet+GMM for BDD dataset, mAP with only 8.+. However, the common mAP for BDD dataset is about 20+. Could u help me...

Shanyaodedanshen commented 3 years ago

Hello, I haven't tried our proposed modeling for the algorithm you mentioned. However, I think that the proposed method can be applied to other networks because most of the output of object detection algorithm is the same. It worked well when I applied it to Faster-RCNN.

  • First, increase the number of output size of both classification and localization heads by considering the number of Gaussian mixture model (GMM) components.
  • Second, modify the loss function as the proposed loss function with the GMM parameters of each head.
  • Lastly, after training is completed, modify the calculation of bounding box coordinates and classification information considering GMM in inference.

Thank you. I have apply your work in RetinaNet. But When I use the raw SSD+GMM for BDD dataset, mAP with only 2.+, which use the same parameters with COCO. And when I use RetinaNet+GMM for BDD dataset, mAP with only 8.+. However, the common mAP for BDD dataset is about 20+. Could u help me...

I use NLL loss and smooth l1 loss meanwhile. Firstly give smooth l1 loss more weight. After the location feature has been trained well, give NLL loss more weight. I also use focal loss, use logits=mu+var*random as the prediction input of focal loss. the result has been normal.

abhigoku10 commented 2 years ago

@Shanyaodedanshen hi i am looking into similar work can you pls share your work ?