Cannot train Detectnet for multiclass

Hi,

I want to train a caffe model to detect car and pedestrian using KITTI dataset. I prepared KITTI dataset according to the session “Downloading and preparing the KITTI data” in this tutorial https://github.com/NVIDIA/DIGITS/blob/v4.0.0/examples/object-detection/README.md

I trained model caffe according to the tutorial https://github.com/NVIDIA-Jetson/jetson-trashformers/wiki/Single-and-Multi-Class-Object-Detection.

Here are my steps with DEGITS:

  1. Create dataset

I filled “DontCare, Car, Pedestrian” in “Custom classes

  1. I created model with these configs:

3.In “Custom Network” tab, I used the network in this link https://raw.githubusercontent.com/NVIDIA/caffe/caffe-0.15/examples/kitti/detectnet_network-2classes.prototxt

  1. In “Pretrained model(s)”, I used bvlc_googlenet.caffemodel from this link https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet

The problem is:

  • mAP and other index of class 1 is always equal to zero
    Please see this image

  • The trained model cannot detect 2 class, as image

Where did I wrong?

Here is my DIGITS info:
DIGITS version: 6.1.1
Caffe version: 0.15.14
Caffe flavor: NVIDIA

Thanks,

@dqthebt … I’m trying to do a similar thing - did you ever get it working? … Thanks!

I’m a bit confused about these lines:

object_class: { src: 1 dst: 0} # cars -> 0
object_class: { src: 8 dst: 1} # pedestrians -> 1

What values would i use for ‘bottles’ and ‘dogs’? Or does it not matter? … And what about if i was using my own class such as ‘rabbits’? Would i have to define rabbits somewhere? (Obviously rabbits is in all the relevant kitti labels).