SqueezeDetMX

This repository is built in Python3.6. With small syntax changes, it could be adapted to Python3.. To my knowledge, those changes would render the repository backwards-compatible with Python2..

This repository contains an implementation of SqueezeDet, a "unified, small, low power fully convolutional neural network for real-time object detection for autonomous driving", in MXNet. Note that this only runs on Python3. It additionally holds utilities for reading and writing compact binary data, on top of deserialization functions for the KITTI dataset.

You can find the original squeezeDet implementation, using Tensorflow, here. Parts of the codebase were taken from the original repository; all such instances have been cited accordingly.

1. Install

(Optional) We recommend setting up a virtual environment.

virtualenv squeezeDetMX --python=python3
source activate squeezeDetMX/bin/activate

Say $SDMX_ROOT is the root of your repository. Navigate to your root repository.

cd $SDMX_ROOT

We need to setup our Python dependencies.

pip install -r requirements.txt

2. Setup KITTI

The first two steps were taken nearly word-for-word from the original README.

First, obtain the KITTI object detection dataset links: images and labels. Start by creating and changing into a directory for KITTI.

mkdir $SDMX_ROOT/data/KITTI
cd $SDMX_ROOT/data/KITTI

Download and then unzip both files.

wget <link to object zip>
unzip data_object_image_2.zip
wget <link to label zip>
unzip data_object_label_2.zip

Train-Val Splitting

Now we need to split the training data into a training set and a validation set. Create a directory to hold both.

mkdir ImageSets
cd ./ImageSets

Then, create a new file trainval.txt, containing indices to all the images in the training data. In our experiments, we randomly split half of indices in trainval.txt into train.txt to form a training set and rest of them into val.txt to form a validation set.

ls ../training/image_2/ | grep ".png" | sed s/.png// > trainval.txt

For your convenience, we provide a script to split the train-val set automatically. Navigate to the data directory's root.

cd $SDMX_ROOT/data

Simply run

python random_split_train_val.py

You should get both train.txt and val.txt under $SQDT_ROOT/data/KITTI/ImageSets.

When above two steps are finished, the structure of $SQDT_ROOT/data/KITTI/ should at least contain:

$SQDT_ROOT/data/KITTI/
                  |->training/
                  |     |-> image_2/00****.png
                  |     L-> label_2/00****.txt
                  |->testing/
                  |     L-> image_2/00****.png
                  L->ImageSets/
                        |-> trainval.txt
                        |-> train.txt
                        L-> val.txt

Converting into RecordIO

This repository additionally contains a conversion script, from KITTI to RecordIO file objects. Navigate to the repository root, and run convert.py.

cd $SDMX_ROOT
python convert.py

If your data was downloaded and setup per the above Data section, the script will require no flags. Otherwise, run python convert.py --help for more information.

3. Train

The repository will support any RecordIO binary, written using the squeezeDetMX.utils.Writer object. By default, the script will look for RecordIO objects in ./data/KITTI. To start running, use

python train.py