A Deep Learning model for classifying on-road entities using CNN with Progressive Resizing. Please refer the report for methodology.
User Guide
-
Creating dataset • Please see sec.2.1 in report for diagrams. • Querying the search engine for the required class labels The required class label can be queried on images.google.com. If majority of the results are satisfactory, proceed. Else a different query can be fired. • Fetching the URL’s of resulting images through JavaScript The TXT file containing the URL’s of the above search results can be downloaded by executing the provided “js_console.js” file, step by step in the JavaScript console of the web browser (under the “developer” menu-bar item of the browser). A file named “urls.txt” gets downloaded at this step. • Downloading the images through python script This “urls.txt” file should be placed in the directory where the “download_images.py” script file is present. In terminal, execute the head command to see the format for providing the “urls.txt” file & destination directory path to this script. Download script should be executed in the following way: Execute – python3 download_images.py –urls urls.txt --output path/to/dir
-
Choosing a predefined dataset • Make sure the class labels are same as (bike,bus,car,pedestrian). • The dataset should be arranged in the hierarchical structure as explained in section 2.2. E.g. data/train/bus data/train/car data/val/pedestrian data/test/bus data/test/bike
• A sample dataset can be downloaded from - https://drive.google.com/file/d/1yAvTpaNdeqVqi7TPo1FHkMltU2wM82ns/view?usp=sharing -
Training • To train the model, place the “train.py” file in the same directory as dataset’s “data” folder. • Execute – python3 train.py
-
Testing • To test the model, place the “test.py” file in the same directory as dataset’s “data” folder. • Make sure the weight files, created by the training script are present in the same directory. • Execute – python3 test.py
Dependencies - imutils argparse requests opencv numpy pandas matplotlib sklearn tensorflow keras Please check versions in requirements.txt