search

qianguih / voxelnet

This is an unofficial inplementation of VoxelNet in TensorFlow.
654 stars 241 forks source link

Training the voxel net for different objects at a time? #90

Closed GuidanceMantras closed 3 years ago

GuidanceMantras commented 3 years ago

Hey,

training the model for different models at a time?

Kindly let me know

Thanks in advance

Manueljohnson063 commented 3 years ago

Hai,I am stuck with this in adding multiple objects in training could you please help how to add multiple objects!!!!

GuidanceMantras commented 3 years ago

@Manueljohnson063

Sorry for the late. I am also stuck there only. I am trying to add multiple objects for training, but so far no success

GuidanceMantras commented 3 years ago

Hi @Manueljohnson063

Here is the solution to add multiple objects to voxelnet. Still it is not efficient one, but it works. So you have to find the cls_loss_pedestrian and reg_loss_pedestrian as well as for cyclist and car for each object individually and add them to find the final cls_loss and reg_loss. Here one has to use hyper parameters to balance the classification loss.

Since most of the frame have more cars than pedestrian, and also most of the frame has more pedestrian than cyclist. It is upto the user how he/she can balance this using hyper parameters.

Another solution would be similar to the above one, but with different loss function using one hot encoding! Since I am doing my master thesis, I do not want to take any risk in trying other solution(efficient one), when existing solution already works.

Hope this helps you!

mulit_classification

sofiaesp commented 3 years ago

@Suri12990 How do you modify model to train for cyclist and pedestrian, and how do you add the loss? I would appreciate your help.

sofiaesp commented 3 years ago

Hi I found that you can modify the selected object in config.py, I have to train a model for each object? If so, when I get the three models (one for each object) how could I add them to find a final loss? @Suri12990

gynaa commented 2 years ago

hi @sofiaesp did you figure it out? i would really appreciate your help