ml-lab / 3dvp

Data-Driven 3D Voxel Patterns for Object Category Recognition

3DVP: Data-Driven 3D Voxel Patterns for Object Category Recognition

Created by Yu Xiang at CVGL at Stanford University.

Introduction

We propose a novel object representation, 3D Voxel Pattern (3DVP), that jointly encodes the key properties of objects including appearance, 3D shape, viewpoint, occlusion and truncation. We discover 3DVPs in a data-driven way, and train a bank of specialized detectors for a dictionary of 3DVPs. The 3DVP detectors are capable of detecting objects with specific visibility patterns and transferring the meta-data from the 3DVPs to the detected objects, such as 2D segmentation mask, 3D pose as well as occlusion or truncation boundaries. The transferred meta-data allows us to infer the occlusion relationship among objects, which in turn provides improved object recognition results.

License

3DVP is released under the MIT License (refer to the LICENSE file for details).

Citing 3DVP

If you find SubCNN useful in your research, please consider citing:

@incollection{xiang2015data,
    author = {Xiang, Yu and Choi, Wongun and Lin, Yuanqing and Savarese, Silvio},
    title = {Data-Driven 3D Voxel Patterns for Object Category Recognition},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    pages = {1903--1911},
    year = {2015}
}

File Organization

1. Geometry: scripts to voxelize 3D CAD models.

   # voxelize 3D CAD models
   cad_train.m

2. KITTI/PASCAL3D: scripts to discover 3DVPs from the KITTI/PASCAL3D+ detection benchmark.

   # create 3D voxel exemplars
   create_annotations.m
   
   # prepare clustering data for 3DVPs
   prepare_clustering_data.m
   
   # clustering to discover 3DVPs
   cluster_3d_occlusion_patterns.m

3. ACF: ACF detectors for 3DVPs

   # training and testing ACF detectors for 3DVPs
   exemplar_dpm_train_and_test_batch_aps.m

4. ContextW: occlusion reasoing with 3DVPs

   # greedy occlusion reasoning with 3DVPs
   greedy_occlusion_reasoning.m