Skip to content

Latest commit

 

History

History
42 lines (40 loc) · 1.84 KB

Kitti.md

File metadata and controls

42 lines (40 loc) · 1.84 KB
Raw

Notes on the Kitti / object detection example

This page is attempted as a guide to accomplishing the object detection example described here in detail: https://github.com/NVIDIA/DIGITS/tree/master/examples/object-detection

We are assuming that you have a VM with Digits in it and your /data folder is passed through to your container..

Pull the GoogleNet weights

cd /data
wget http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel

Clone the DIGITS

cd /data
git clone https://github.com/NVIDIA/DIGITS.git

Get the KITTI data

Go to the appropriate directort

cd /data/DIGITS/examples/object-detection

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

Once the data is downloaded, run

./prepare_kitti_data.py

You can now follow the rest of the guide, just remember that your datasets are located under

/data/DIGITS/examples/object-detection/kitti-data

And your pre-trained GoogleNet weights are located under

/data/bvlc_googlenet.caffemodel

Notes on training

If you run the training as suggested in Nvidia github, you likely will not get to the results shown there (mAP over 50%) on the first try. The set up does not appear to be very stable, and you are more likely to see something like this: Fig1 In this case, the mAP suddenly crashed at epoch 8 and we aborted the run. It would be possible to continue training, and we are likely end up with something like this: Fig2 Here, the mAP made a recovery, but not a full one. However, another idea is to recover from the checkpoint just before the crash, and "roll the dice" again. Odds are, you will have better luck -- as we had here: Fig3 Do give this some thought. This is the essential idea behind transfer learning.