Skip to content

Latest commit

 

History

History
52 lines (30 loc) · 2.24 KB

README.md

File metadata and controls

52 lines (30 loc) · 2.24 KB

Depth-Segnet

Segnet using RGB-D images

Segmentation done using Depth as an additional parameter.
Depth is converted into three fields : horizontal disparity, height above ground, and the angle the pixel’s local surface
normal makes with the inferred gravity direction.
Idea taken from paper Depth-RCNN.

This network takes as input a RGB+HHA image[6-channels].
Script to make a lmdb from RGB + HHA + Segmented mask file is provided.
You can also download the converted lmdb from:
Data(RGB+HHA+ Segmented Masks)link
The modified prototxts are included in models as (segnet_depth_*).
NYUdv2 dataset was used for training. Follow the rest of instructions from original Segnet.

Modified scripts can be found in datatools folder.

Setup

  1. Clone this repository: git clone https://github.com/hari-sikchi/DepthSegnet.git
  2. Initialize all submodules: git submodule update --init --recursive
  3. Download all the data from the link given below or create your own data in the format:
    Data/data_lmdb(Containing images of 6 channel)
    Data/labels(Containing segmentation masks for each channel)
  4. Start training and see for yourself!

Examples

Segmentation results in cluttered scenes from NYUDv2 dataset [Added Filter output visualization]. alt text

Final trained weights can be found here. Link

References

SegNet: http://mi.eng.cam.ac.uk/projects/segnet/tutorial.html