By Mousa Arraf & Chen Chen Katz.

This repository is part of a final project in the Technion's course 046211 - Deep Learning

In this work we try to boost the accuracyof basic CNNs and CNN with an LSTM layer using an estimation of the depth of the objects in RGB pictures.

cifar-10 RGB format which is available at : https://www.cs.toronto.edu/~kriz/cifar.html

Object-RGBD which is available at : https://rgbd-dataset.cs.washington.edu/dataset.html

1. git clone https://github.com/arrafmousa/Deep_final_project after cloning in the repository please run :

   conda env create -f environment.yml
   conda activate depth_estimation
   

2. download the rgb dataset from: http://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
3. run

   # create cifar-10 RGBD dataset:
   python create_cifar10_depth_dataset.py --cifar10_dir [the rgb dataset you downloaded] --cifar10_RGBD_output_dir [rgbd output dir]
   
   # get the accuracy improvment ratio of diffrent datasets sizes in case of using estimated depth images:
   python rgbd_cifar10_dataset_classifier.py --cifar10_rgbd_dir  [the created rgbd dataset path] --train_with_all_dataset_sizes
   
   # get the accuracy improvment ration of diffrent small datasets sizes in case of using estimated depth images: 
   python rgbd_cifar10_dataset_classifier.py --cifar10_rgbd_dir  [the created rgbd dataset path] --train_with_small_dataset_sizes
   
   # test the performance of the diffrent architectures
   # where you can pass True or False as is_lstm or is_depth to train diffrent models.
   python conv_neural_network.py --is_lstm=is_depth --is_depth=is_depth --cifar_dir=[the created rgbd dataset path]
   

here we can see that using estimated depth we were able to boost the performance of our baseline models.

we also experimented with diffrent sizes of datasets the accuracy boost is as follows

As expected, the biggest improvement in performance can be achieved when using small datasets