HOU Yuenan
HOU Yuenan
You should have a look on the variable name of ry's pre-trained model. Then, you should use the same variable name and this problem should be fixed.
The following code works for me where I load the weight of 2D ResNet-50 model into 3D ResNet-50 model: `with tf.Session() as sess:` ` # build 3D ResNet-50` `saver =...
The variable name is stored in `saver._var_list`.
The following code works for me where I load the weight of 2D ResNet-50 model into 3D ResNet-50 model: `with tf.Session() as sess:` ` # build 3D ResNet-50` `saver =...
@SongleChen2015, the following URL stores the model of 3d resnet-50, please check. https://drive.google.com/open?id=1WkNtPNqf7-O1-RvXj4auWTH6OYqA2QKG
You can just modify the size of the input images and the model definition. Then, just run the codes on your own dataset.
Currently, the performance of adding the sigmoid layer and using BCE loss is bad (mse loss on the testing set is around 7). I will share the results here if...
@ohosh @huixiancheng I have trained the CENet which can achieve 65 mIoU on SemanticKITTI test set with 64x512 input resolution. The trained model will be put in https://github.com/cardwing/Codes-for-PVKD.
More powerful range-image-based models built on the awesome CENet will also be put in that repo (70+ mIoU on SemanticKITTI test set).
The CENet with 64x512 input resolution has been uploaded to https://github.com/cardwing/Codes-for-PVKD. The reproduced performance (67.6% mIoU) is much higher than the reported value on SemanticKITTI test set.