Boat detection

This project focuses on detecting boats with different points of view, sizes, shapes, types, background elements and illumination conditions.
The main project directory is 'src_cpp' where a CMakeLists.txt file is used to build the application. The main.cpp file implements the proposed approach on a given input image, while in the subdirectory 'utils' there are all the files needed in order to do so. In the folder 'src_python' there is instead a jupyter notebook which shows all the pipeline followed for traning the CNN classifier that will be used in the C++ application. To transfer the model from TensorFlow to OpenCV dnn module, directory 'my_model' and file frozen_graph.pb are needed. Especially the last one is the file used in OpenCV for loading the net configuration.

Brief description

The boat detection algorithm here implemented uses selective search as region proposal method to find possible locations of boats in an image, moreover a CNN made by scratch is used to classify the bounding boxes mentioned before. Then a pre and post processing phases are also added.

The main steps are:
1. Read an input image;
2. Apply a median and bilateral filter;
3. Use Selective Search ("fast mode") to the filtered image for generate regions proposal;
4. Discard all regions that have the ratio between height and width (or vice versa) bellow a given threshold;
5. Each region proposal represent a patch from the original image that will be resized and classified thanks a CNN;
6. Apply Non-Maximum Suppression (NMS);

The final results are quite nice for some input images like:

while in photo with a complex background, the detection is more challenging:

For further details look at the report: 'report/report_boat_detection.pdf'

Basic Usage

src_cpp/main.cpp

In the first modality the application produces in output the image with green bounding boxes for all regions with boats. Above each box a score is present and represent the probability of that region to be a boat. As command line argument, give only the image path:

image_path

For example: 01.jpg

In the second modality in addition of what happens for the first one, in output is also produces the comaprison with the ground truth (Note: the assumption is that the ground truth is stored with same format as in dataset, for example boat:264;371;342;362;). As command line arguments, give the image path and ground truth path (.txt file):

image_path ground_truth_path

For example: 01.jpg 01.txt

Note: Depending on the image, the computation may require some moments.