Absolutely zero mAP, precision, recall while trying to train a pre-trained caffe model using KITTI o...

AI & Data Science Deep Learning (Training & Inference) DIGITS (Locked)
1

We are trying to train a pre-trained .caffemodel on KITTI object detection dataset.

Our dataset files include: http://www.cvlibs.net/datasets/kitti/eval_object.php

  1. Left color images of object data set (12 GB)
  2. Training labels of object dataset(5 MB)
  3. Object Development Kit(Shows 1 MB but is actually 64 KB)

This is the link of the tutorial that we have followed:

https://github.com/NVIDIA/DIGITS/tree/master/examples/object-detection#downloading-and-preparing-the-kitti-data

We are getting absolutely zero mAP, precision, recall values even after 30 or 50 epochs.

All the configurations as exactly the same as mentioned in the GitHub link above.

Can anyone guide us in this, please?

2

Did you put the googlenet as pre-trained network? Otherwise the front part of your network would be empty and it requires very long training to actually get something on mAP, precision, recall.

3

I know this is quite old but I was having the same issue. What I had to do was change the Blob Format box on model creation page from NVCaffe to Compatible. Once I did that and started my training after about 100 or so epochs I started to see values other than 0 for mAP, precision, and recall.