This repository contains the code used to train and test the network described in the paper Scene Structure Inference through Scene Map Estimation.
Note: The current data available is the data used to generate Figure 8 in the paper, i.e. the overly dense and cluttered scenes. This means the performance you will get is close to what you see in Table 2, not the one in Table 1. The data for Table 1 will be released soon.
All code, except for where specified, written by Moos Hueting.
Currently, we release the model files for the depth, semantics and scene map estimation networks, as well as a set of data generated using the rendering pipeline described in the paper. We will be releasing the rendering code in the future as well.
This software is released under the MIT License (refer to the LICENSE file for details).
- Torch
- [CuDNN]
- Clone the repository
git clone https://github.com/mxh/scenemap.git
- Download the model and data files
# The directory in which the repo has been cloned is $SM_ROOT
cd $SM_ROOT
./download_models.sh
./download_data.sh
Each network (depth, semantics and scenemap) can be run using their respective shell scripts:
# depth
./run_depth.sh
# semantics
./run_semantics.sh
# scene map
./run_map.sh
The output shows performance, and the output can be judged qualitatively in test_*.png.