No description, website, or topics provided.
C++ Matlab Cuda C
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data
lib
src
util
.gitignore
LICENSE
README.md
demo.m
eod_compile.m
eod_default_param.m
startup.m

README.md

Enriching Object Detection

This is a code repository for my CVPR 15 paper, Enriching Object Detection with 2D-3D registration and continuous viewpoint estimation.

@InProceedings{Choy_2015_CVPR,
  author = {Bongsoo Choy, Christopher and Stark, Michael and Corbett-Davies, Sam and Savarese, Silvio},
  title = {Enriching Object Detection With 2D-3D Registration and Continuous Viewpoint Estimation},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2015}
}

Pipeline

Dependencies

There are two modes of operation. EOD works as a standalone detector and also as a detection augmentation pipeline. The demo code contains both versions of mode.

Installation

The OSGRenderer requires OpenSceneGraph

To install on Mac OS X, type

brew install open-scene-graph

To install on Linux, type

sudo apt-get install openscenegraph

Next, install the OSG and this code repository and install by typing

git clone https://github.com/chrischoy/EnrichingObjectDetection
git clone https://github.com/chrischoy/OSGRenderer
cd OSGRenderer
matlab -r compile
cd ../EnrichingObjectDetection
matlab -r 'eod_compile'

Demo

First, the demo code will generate NZ-WHO detectors. The detectors are generated on the fly without training.

NZ-WHO detectors

Next, using the detectors, we can detect objects in the image. However, we can feed in the detection results from any other off-the-shelf detectors.

Detections

Using our detectors, we can get the detection results.

detection result

Finally, we finely tune the detection result using MCMC stage.

Detections

After the tuning,

tuning result 1

You can observe that the score increased from 96 to 111 and the overlap (Intersection over Union) between the prediction and the ground truth bounding box increased from 0.78 to 0.8 as well. Notice that the car slightly tilted forward to match the image exactly.

For more results, please visit https://github.com/chrischoy/EnrichObjectDetectionResults

Issues

Mac OS Matlab can't find a compiler

Follow the instruction and setup the xcode compiler