search

alirezazareian / vspnet

Code for the CVPR 2020 oral paper: Weakly Supervised Visual Semantic Parsing
https://arxiv.org/abs/2001.02359
35 stars 6 forks source link

Pretrained Model on OpenImage #8

Closed zhaohengz closed 4 years ago

zhaohengz commented 4 years ago

Hi,

Thanks for the wonderful work!

May I ask where can I find the pretrained model that you guys took for feature extractor and proposal generator?

Thanks a lot for your help.

Best, Zhaoheng

alirezazareian commented 4 years ago

I used this tool to extract the proposals and features. More specifically, I used the checkpoint named faster_rcnn_inception_resnet_v2_atrous_lowproposals_oidv2 from here for the weakly supervised experiments, and faster_rcnn_inception_resnet_v2_atrous_oidv2 for fully supervised. The only difference is the number of proposals that they generate.

zhaohengz commented 4 years ago

Thanks a lot for helping out!

Best, Zhaoheng

zhaohengz commented 4 years ago

Hi,

Sorry for bothering again.

I am trying to extract features and bounding boxes from the model. On which level do the boxes come from, the proposal level or after post-processing?

Thanks for helping out.

Best, Zhaoheng

yekeren commented 3 years ago

I've got the same question. Wish the authors to respond.

My guess is the average pooling(called in https://github.com/tensorflow/models/blob/72a31e9e5c06a663a8ee7ff7cc2a41e20557c9e5/research/object_detection/predictors/heads/class_head.py#L85) of the predictions['box_classifier_features'] (aka, SecondStageFeatureExtractor/InceptionResnetV2/Conv2d_7b_1x1, shape=num_proposalsx8x8x1536)(https://github.com/tensorflow/models/blob/72a31e9e5c06a663a8ee7ff7cc2a41e20557c9e5/research/object_detection/meta_architectures/faster_rcnn_meta_arch.py#L1061).

I am not sure which boxes are used (refined or not refined).