Question, how to run without camera
Closed this issue · 10 comments
Hi!
Thank you for sharing your excellent work. In the paper there is a row in Table III where you use BEV only features (RPN BEV Only). I am interested in using this network, is there a configuration file available for this?
Also, would it be possible for you to share the trained models?
We did not make this configurable however modifying the RPN to use BEV only is fairly straightforward. Inside the rpn_model.py, you need to disable image and BEV fusion (see code proposal_roi_fusion and instead set the rpn_fusion_out to bev_proposal_rois features directly. You can then commend out the generation of img_proposal_rois as it won't be needed. Note that this mode only disables BEV for RPN, the second stage still uses the image features.
With regards to sharing trained models, since the network trains relatively fast and there is no pre-training required, we decided not to publish the trained models.
Hope this helps.
@melfm Thanks for answering, I'm wondering whether we can use BEV feature for the second stage so we only need the BEV input for the model? is that possible or it may harm much performance?
Thanks!
You could try using the BEV features only for second stage too, however the performance is expected to drop (specially for detecting smaller objects like pedestrian and cyclists).
@melfm Thanks for the answer. I was wondering if we can do this by setting path_drop_probabilities to [0, 1].
Edit: Changing path_drop_probabilities would affect both the first and second stage if my understanding is correct.
Path_drop gets disabled during evaluation and testing, so it's not the best way to disable the image.
@melfm That makes sense. It would involve more changes and it is not really computationally efficient neither. Thanks for the clarification.
@melfm Thanks for the answer. I was wondering if there is any lidar-only detection result on Kitti that we can refer to make sure whether our modification is correct?
@LouieYang please check our paper for RPN with BEV only results on the validation set.
@LouieYang Sorry for bothering you, can I ask for the BEV only version.
I'm trying to modify the code but It seems too complicated for me.
@LouieYang Sorry for bothering you, can I ask for the BEV only version.