The obdes from soroosh-n

ObDES: Object Detection and depth EStimation

Sharif University of Technology
Deep Learning course project - Fall 1400
Supervisors: Prof. Emad Fatemizadeh - Behrad Ahmadpour

📝 Table of Contents

About project
Getting Started
Some results
Contributors

About project:

In this project, we created a PyQt5-based application that will take a photo as input and produce the same photo with bounding boxes around the objects. For each of those bounding boxes, there will be a label of the object's category. With what confidence? And how far is that object from the camera?
(On a scale of 0 to 1 [ 0: so close, 1: far, far away])

Getting Started:

On this project, we have three main files:

ObDES_colab_notebook.ipynb: This is the project's first and most complete file. It runs on the Google Colab, and it's not related to the GUI section of the project. But it will create one of the most important files, which is necessary to make the GUI section able to work: Estimation model w8s (weights).
This notebook file consists of 3 main parts. In the first part, we use the NYUv2 dataset to produce enough pictures to learn the estimation model, which is an autoencoder network.
In the second part of the file, we will train the estimation model by using the pictures created in the previous part. Then we save the created model as estiamtion_model.h5 in the resources directory.
In the third part of the project, we will use the produced pictures and the learned model in collaboration with YOLOv3 pre-trained model to finalize our project. First of all, we will also download and save the YOLOv3 necessary files (coco.names, yolov3.cfg and yolov3weights) in the resources directory. And we directly create the joint output of the whole network. We will use the bounding box to identify which pixels should be considered as the object pixels. Then we will take an average on those pixels, and the final result will be regarded as the object's distance from the camera.
ObDES.py: This is the one-run file of almost the whole project without the GUI features. In this file, we will use the results of the first two parts of the Google Colab notebook to run the YOLO model and take the results directly. For this purpose, all four files of the resources directory should be downloaded before the main processes. This downloading process will be done automatically. Just give this file a good internet connection and wait for about 640 MB of files to be downloaded (240 MB is for the YOLOv3 weights, and almost the rest is for the estimation_model.zip file).
ObDES_GUI.py: The automatic downloading procedure is like the last part. But first, you can check what files are missing. This file is nothing but a PyQt-based GUI of the ObDES.py file.
These instructions will get you a copy of the project on your local machine for development and testing purposes. See deployment for notes on deploying the project on a live system.

Prerequisites

We need some libraries to be installed on the environment first. We used a python 3.8 library for the project.

General libs:
  os,
  sys,
  time,
  numpy,
  shutil,
  zipfile,
  traceback,
  progressbar,
  urllib.request,
  matplotlib.pyplot,

To be installed:
  cv2,
  pillow,
  tensorflow + keras,
  PyQt5 (QtCore, QtGui, QtWidgets),
...

Some packages version:

Versions:
OpenCV      => 4.5.5
PyQt        => 5.9.7
Tensorflow  => 2.7.0

Some results:

Our model works fine with almost every picture. But it will be much better if we train it with more images. We train the estimator model only with 1449 pictures of the NYUv2 dataset and under 60 epochs. This dataset count is absolutely nothing against the coco dataset, which has about 338 k pictures. By the way, our current model has excellent capabilities. We bring some results here as examples: