This repository works upon training a neural network for estimating depth from monocular images that can later be used for 3D reconstruction of the scene.
The companion blog post can be found here.
The work is based upon this paper.
13 GB of Data uploaded by Raghav Prabhakar can be found here
The dataset contains RGB images and their corresponding depth maps encoded in CARLA's RGB format.
- Download and place the depth and RGB images in their corresponding folders inside the 'data' directory.
- Install the dependencies used in the project.
- Run
python train.pyto start training.
Full video can be found here
How to create/ visualize this inverse projection image/ video [See here]
The network was collaboratively trained by Raghav Prabhakar, Chirag Goel, Mandeep and me on google colab for 20 Hrs.
Pretrained model (With DepthNorm Training) can be found here.
- Add model training scripts.
- Add data collection scripts for CARLA.
- Add easy hyperparameter tuning through a seperate hyperparameter file that can easily be edited.
- Add 3D reconstruction Code.
- Add model evaluation code.
- Train a model without DepthNorm.
- Add more exception handling .
- Add CLI for easy argument passing.