search

xiaohaoChen / rrc_detection

Accurate Single Stage Detector Using Recurrent Rolling Convolution
Other
361 stars 136 forks source link

How about the performance of multi-class detection ? #15

Open catsdogone opened 7 years ago

catsdogone commented 7 years ago

@xiaohaoChen @jimmy-ren

Hi! Thanks for your code and paper.

It seems that you code is designed for one class training.

If I want to train a multi-class model, a model that can detect car, person and bicycle at the same time, Should I change the loss function? And do you think the multi-class model will has good performance as described in your paper?

Thank you very much.

jimmy-ren commented 7 years ago

What we found during the experiments on KITTi was using a single model for car and a single model for both pedestrian and cyclist gave the best results. Many other papers in the literature seem to did the same. This may due to the limitation (limited groundtruth) of the KITTI dataset. In theory, if you have enough data, train a model to detect all the categories shall also give good results.

zhanghanduo commented 5 years ago

@jimmy-ren @xiaohaoChen Is it possible to provide the trained model weights for pedestrian and cyclist classes ? Thank you

GreyCoder25 commented 2 years ago

@jimmy-ren @xiaohaoChen Is it possible to provide the trained model weights for pedestrian and cyclist classes ? Thank you