How to test on pedestrian and cyclist class?
gujiaqivadin opened this issue · 5 comments
Hello, zhixinwang!
Thanks for your great work on frustum convnet. I found that there is only configs code for car training, if I want to test the performance on people class, what do I need to change in scripts or configs files.
Hello @zhixinwang , i too have a similar question. Does your git repo cover implementation for cyclists and pedestrians too? how to train and test for those objects?
Sorry, current release version does not include the implementation for pedestrains and cyclists.
If you want to train pedestrians and cyclists, you can follow the same step during training the car model, but you need to change some config settings, such as 'CAR_ONLY=False,PEOPLE_ONLY=True', 'IOU_THRESH=0.5', and the STRIDE and HEIGHT_HALF parameters according to our paper.
The above config changes mentioned for performing the train and validation on Pedestrian and Cyclist classes, are these changes enough or do we need to change something in code?
And we need to perform the above changes in "det_sample.yaml" and "refine_car.yaml" files? or any other changes required?
I think changes of the config in *.yaml is enough.
The most import settings you need to change are STRIDE and HEIGHT_HALF. For reference, in the first stage, STRIDE: (0.1, 0.2, 0.4, 0.8), HEIGHT_HALF: (0.1, 0.2, 0.4, 0.8), and refine stage is STRIDE: (0.05, 0.1, 0.2, 0.4), HEIGHT_HALF: (0.05, 0.1, 0.2, 0.4) for people category.
Thank you for the response.
I do have another question to clarify!
The variable MAX_DEPTH - what is it used for? is it to limit the range of LiDAR data or it has some other significance?