This was done as part of my B.Tech. project in Semester 6 under the guidance of Prof. Rajeev Srivastava.

Objective

Find objects in images from any of the 20 PASCAL VOC classes!

Run

Open terminal in project path. Run jupyter-notebook. Open Detection.ipynb and run it.

Code Explanation

1. Region Representation and Extraction

• Regions which are extracted from images using selective search: A region is represented as a rectangular box (y0,x0,y1,x1) where (y0,x0) are top left coordinates and (y1,x1) are bottom right coordinates with y denoting row and x denoting column in image. images_regns.npy: Regions extracted from Training images in PASCAL VOC. test_regns.npy: Regions extracted from Testing images in PASCAL VOC.

• Regions which are given as annotation in PASCAL VOC dataset: A region is represented as a rectangular box (x0,y0,x1,y1) where (x0,y0) are top left coordinates and (x1,y1) are bottom right coordinates with y denoting row and x denoting column in image. ground_truth.npy: Regions extracted from Training annotation of images in PASCAL VOC. test_ground_truth.npy: Regions extracted from Testing annotation of images in PASCAL VOC.

2. Feature Extraction

Postives and negatives (HoG features) for each class are extracted and stored in cats_vs_dogs_positives.npy and cats_vs_dogs_negatives.npy respectively.

3. Training (in Detection.ipynb)

The training is done using extracted features(after applying pca on them). Tuning finds best gamma,C and components(for pca).

4. Testing (in Detection.ipynb)

• Using the test image annotation for boxes, test features are found. Classification report is made.
• Using test images, detection boxes are printed to files which are name as comp3_det_className.
• Detection files can be used to generate precision recall graps after submission to PASCAL VOC.