3DOP -- 源代码结构

通过3DOP项目的主页，我们可以找到通往多伦多大学的项目源代码的链接。不知是处于什么目的，该代码是以压缩包而不是github的方式给出的。

通过下载之后，我们可以看到指导说明为：

3DOP_code_cuDNNv3 and v5

由于cuDNN 3版本和cuDNN5的版本上有一定实现差异，因此该代码兼容了上述两个版本。下面仅针对cuDNN v5 来进行代码分析。

• ## 3D Object Proposals (3DOP)
  This is an implementation of the algorithms described in

3D Object Proposals for Accurate Object Class Detection.
Xiaozhi Chen*, Kaustav Kunku*, Yukun Zhu, Andrew Berneshawi, Huimin Ma, Sanja Fidler and Raquel Urtasun. NIPS, 2015.

Precomputed Results

We provide following precomputed models and results on our project page: http://www.cs.toronto.edu/objprop3d/
- Train-val split of KITTI training set
- Disparity computed by SPS-stereo for KITTI training and testing sets
- Road planes for KITTI training and testing sets
- 2D/3D proposals for KITTI training set
- CNN models trained with 3DOP on KITTI training set
- 2D/3D detection results using 3DOP on KITTI validation set

Compute Proposals

Please refer to ./3DOP/demo_3dop.m for training/testing/evaluation of the proposals.

Object Detection Networks

Detection code is under ./frcn-kitti/
This implementation is built on Fast R-CNN (https://github.com/rbgirshick/fast-rcnn).

1. Build Caffe

   Please follow Fast R-CNN (https://github.com/rbgirshick/fast-rcnn) to compile caffe and pycaffe.

2. Build the Cython modules

   • cd ./frcn-kitti/lib
   • make

3. Prepare data

   • KITTI images are supposed to be placed unber ./frcn-kitti/data/kitti/object/
   • 3D object proposals are supposed to be placed under ./frcn-kitti/data/proposals/

4. Download pre-trained ImageNet models

   • cd ./frcn-kitti
   • ./data/scripts/fetch_imagenet_models.sh

5. Compile KITTI evaluation code

   • cd ./frcn-kitti/kitti/eval/cpp
   • g++ evaluate_object.cpp -o evaluate_object

6. Training & testing

Please refer to ./frcn-kitti/demo.sh for training/testing of the detection networks.

ImageSets

ImageSets就是将KITTI中的7517张图片，随机均匀地划分成两部分。一部分用作test，另外一部分用作validate。两部分的图片数量大致在3750张左右。

其他部分的内容将在下载完成之后进行相关的测试。

参考文献

1. Supplementary Material: 3D Object Proposals for Accurate Object Class Detection
2. 3D Object Proposals using Stereo Imagery for Accurate Object Class Detection