search

qianguih / voxelnet

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

For pedestrian/cyclist training #87

Open arxidinakbar opened 4 years ago

arxidinakbar commented 4 years ago

Hi, hoping for some discussion about parametes setting for pedestrian/cyclist training. I have trying to train for pedestrian&cyclist detection task, and follwed the original paper set the stride to 1 in first block, and feature map size from RPN would be [batch, 200, 240, 128], but cfg.FEATURE_HEIGHT & cfg.FEATURE_WIDTH in config.py is half size of it, because of the cfg.FEATURE_RATIO=2， SO, should we change cfg.FEATURE_RATIO=1 to make network available for training??

I know this site no that active, but still hope someone to exchange the idea, thanks.

charlescho64 commented 3 years ago

Hello, arxidinakbar

Did you solve this issue?

hopefully I also train multi objects, car, pedestrian, and cyclist, but I don't know what should i do...

if you made it, I hope you give me information. plz.

GuidanceMantras commented 3 years ago

@charlescho64 @arxidinakbar

you can refer to this link for hint!

hope it helps both of you.

charlescho64 commented 3 years ago

@Suri12990

Thanks for your advise.

I try to train with kitti datasets and this git code. your word is very good but I don't know which file do i modify. I am sorry I hope you give me more information.

BR

GuidanceMantras commented 3 years ago

Hey @charlescho64

Give me sometime, I will put the whole model for multi label classification in my repo in few day. I am training the model for multi-label classification. Once I got satisfactory results. I will upload the results and model in Github.

sofiaesp commented 3 years ago

Hi, im very interested on th model for multi label classification, is this based on voxelnet? @Suri12990

charlescho64 commented 3 years ago

@Suri12990 Hello, Any progress? I am interesting in your job... If you have any progress, Plz, share it with us...

mariya12290 commented 3 years ago

@charlescho64 Sorry for the delay, My approach is worked and also I am getting better results, more than 50 mAP for both Pedestrain and Cyclist. If I consider just one Class, I am getting almost 70 mAP. Since my work is only for Pedestrian and Cyclist, I consider only those classes.

But I need some time to put the model in Github with all the documentation, Since it is my master thesis, I am using two data sets and annotation of custom data set, finally my thesis report. It takes lots of time to all this things.

By the end of this month I will put in the github.

Never mind.

mariya12290 commented 3 years ago

you can find it here now https://github.com/mariya12290/VoxelNet_MultiClass/tree/master