SqueezeDetTRT

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 TensorRT and CUDA for inference acceleration.

You can find the original squeezeDet implementation, using Tensorflow, here.

Prerequisites

This requires Ubuntu 16 (Xenial Xeres) or later to get the proper libraries work. It should work for other distributions, but I haven't tested them yet. From a terminal, execute the following:

sudo apt-get install build-essential python perl

You also have to install OpenCV 3.3 or higher according to its website here.

And you also have to install CUDA 8.0 and TensorRT 3.0 libraries, according to their websites CUDA Toolkit 8.0 and TensorRT. Also remember to put nvcc (usually in /usr/local/cuda/bin) in environment variable PATH.

Build

Use make from a terminal in this folder to compile executable binary.

Usage

After compilation, use ./sqdtrt -h to learn the usage for this program, as follows.

Usage: sqdtrt [options] IMAGE_DIR RESULT_DIR
Apply SqueezeDet detection algorithm to images in IMAGE_DIR.
Print detection results to one text file per image in RESULT_DIR using KITTI dataset format.

Options:
       -e, --eval-list=EVAL_LIST_FILE          Provide an evaluation list file which contains
                                               the image names (without extension names)
                                               in IMAGE_DIR to be evaluated.
       -v, --video=VIDEO_FILE                  Detect a video file and play detected video
                                               in a new window. IMAGE_DIR and RESULT_DIR
                                               are not needed.
       -b, --bbox-dir=BBOX_DIR                 Draw bounding boxes in images or video and
                                               save them in BBOX_DIR.
       -x, --x-shift=X_SHIFT                   Shift all bboxes downward X_SHIFT pixels.
       -y, --y-shift=Y_SHIFT                   Shift all bboxes rightward Y_SHIFT pixels.
       -h, --help                              Print this help and exit.

Demo

There are two demoes for image and video detections below. The image and video for demoes are located in data/example.

Image detection

The following command will detect an image in data/example and print bounding boxes using KITTI format in data/result/sample.txt

./sqdtrt -e data/example/val.txt data/example data/result

Video detection

The following command will detect a video in data/example and play it with bounding boxes in a new window. There is some slight dislocation due to the image resize operation (maybe?), so we have to use the -x '-20' -y '-20' arguments to fix it.

./sqdtrt -v data/example/20110926.avi -x '-20' -y '-20'

zhaozhixu / SqueezeDetTRT