This repository contains the code for a monocular depth estimation project, created as part of a recruitment task assigned by Cendas Gmbh. The project utilizes a U-Net architecture for monocular depth estimation.

The goal of this project is to implement a monocular depth estimation model using deep learning techniques. First, a simple baseline model is trained with an encoder-decoder architecture. Then, the U-Net architecture is employed for its effectiveness in image-to-image regression tasks. This repository contains scripts for training, testing, and visualizing the output for the monocular depth estimation task using a U-Net architecture, leveraging the NYU Depth V2 dataset.

• scripts/: Directory containing Python scripts.
  • baseline_model.py: Python script for creating simple encoder-decoder architecture with NYU Depth V2 dataset available in TFDS.
  • train.py: Python script for training the monocular depth estimation model with U-Net and using a subset of the NYU Depth V2 dataset downloaded from kaggle.
  • evaluate.py: Python script for testing the trained model on new data.
  • utils.py: Utility functions used in the project.
  • loss.py: Implementing custom loss which is a combination of SSIM, L1, and Edge Loss.
  • model.py: Implementation of the U-Net model for depth estimation.
  • visualize.py: Script for visualizing the results of the depth estimation.
• Literature/: Directory containing published papers for bibliographic research work.
• outputs/: Directory containing the depth images predicted by the trained models.
• plots/: Directory containing the loss curve for training the U-Net model.
• Assessment Report: This file contains a brief explanation about each task assigned for the project.

git clone https://github.com/neelsen1994/depth_estim_cendas.git
cd depth_estim_cendas
pip install -r requirements.txt

Download the dataset from here, create a data/ directory in the project root directory, and store it there so that it can be easily accessed from the code. The dataset directory structure should be the following:

/depth_estim_cendas
  ├── data
  │   ├── nyu2_train
  │   └── nyu2_test
  │   └── nyu2_train.csv
  │   └── nyu2_test.csv

To train the model, use the train.py script and run the command from the project root directory. You can customize the training parameters by modifying the script.

python ./scripts/train.py

The trained model is available for download below. Create a model/ directory in the project root directory, and store the downloaded model there so that it can be easily accessed from the evaluation script.

Baseline model: Download Model

U-Net Model: Download Model

For testing the trained model on new data, use the evaluate.py script and run the command from the project root directory.

python ./scripts/evaluate.py

The visualize.py script provides tools for visualizing the results of the depth estimation on sample test images.

python ./scripts/visualize.py

The project was inspired as part of the assessment for the hiring process at Cendas Gmbh. I would like to express my gratitude to Alex Zinelis, Klari and Cendas Gmbh for the opportunity to work on this project.

If you have any questions, please do not hesitate to contact me

• Debayan Sen ✉️ debayan.sen1994@gmail.com