Add files via upload

Author: nipunmanral

README.md

@@ -9,6 +9,7 @@ Python Version = 3.6
 OpenCV Version = 4.1.1
 
 #Code Execution
+
 **Selective Search**
 To run the code, download it and execute using the following command in terminal/command prompt:
 
@@ -20,6 +21,7 @@ To run the code, download it and execute using the following command in terminal
 
 Eg: python selective_search.py ./JPEGImages/1.jpg ./Annotations/1.xml color
 
+
 **EdgeBoxes**
 To run the code, download it and execute using the following command in terminal/command prompt:
 
@@ -31,7 +33,7 @@ To run the code, download it and execute using the following command in terminal
 Eg: python edgeboxes.py ./JPEGImages/1.jpg ./Annotations/1.xml
 
 ## Data
-The sample color images are present in '/JPEGImages/' folder and their corresponding ground truth boxes are present in '/Annotations/' folder. An example of the bounding box coordinates (saved in .xml file) is in the following format:
+The sample color images are present in `/JPEGImages/` folder and their corresponding ground truth boxes are present in `/Annotations/` folder. An example of the bounding box coordinates (saved in .xml file) is in the following format:
 
 <object>
   <name>aeroplane</name>
@@ -69,7 +71,7 @@ The steps followed in implementing the code are as follows:
 14) Find those bounding box that have the maximum/best overlap for each ground truth box and display them.
 15) Calculate the recall for each image.
 
-###EdgeBoxes
+### EdgeBoxes
 The Edge Boxes algorithm uses edges to detect objects. The overlying idea of this algorithm is that if a bounding box contains a large number of entire contours, then the bounding box has a high probability to contain an object. More information about EdgeBoxes can be found in the paper (https://www.researchgate.net/publication/319770284_Edge_Boxes_Locating_Object_Proposals_from_Edges)
 
 The steps followed in implementing the code are as follows:
@@ -90,27 +92,27 @@ The steps followed in implementing the code are as follows:
 15) Experiment with different values of alpha and beta.
 
 ## Evaluation Metric
-**Intersection over Union**
+### Intersection over Union
 We evaluate the proposed bounding boxes generated by Selective Search and EdgeBoxes algorithms using Intersection over Union (IoU) metric which is defined as:
 
   Intersection over Union (IoU) = Area of overlap / Area of union
 
 Here, Area of overlap is the area of intersection between the ground truth box and the proposed bounding box. Area of Union is the total area encompassed by both the ground truth box and the proposed bounding box. If the ground truth box and the proposed bounding box do not overlap, then the IoU is 0.
 
-**Recall**
+### Recall
 We calculate the “recall” of ground truth bounding boxes by computing what fraction of ground truth boxes are overlapped with at least one proposal box with Intersection over Union (IoU) > 0.5.
 
 ## Analysis
 My analysis on these algorithms are listed below for one of the images (Image 5 in the *Results.pdf* file)
 
-**Selective Search:**
+### Selective Search:
 The test image shown in Appendix A - Image 5 has been annotated with 13 ground truth boxes. The results obtained using color strategy only have a total of 334 proposed bounding boxes. The number of proposal boxes found to have an IoU greater than 0.5 with any of the ground truth boxes is six. Out of these six proposal boxes, only four ground truth boxes have an overlap with atleast one proposal box. The recall for this image using only color strategy is 0.307.
 
 The results obtained using all strategies i.e. color, fill, texture and size contain a total of 354 proposed bounding boxes. The number of proposal boxes found to have an IoU greater than 0.5 with any of the ground truth boxes is eight. Out of these eight proposal boxes, only seven ground truth boxes have an overlap with aleast one proposal box. The recall for this image using all strategies is 0.5384.
 
 In general, Selective Search using all strategies had a better performance (better recall) when compared to just the color strategy. In our particular image (Image 5) that we considered, we can notice that the image has a variety of colors. Since our code utilizes only the top 100 proposal boxes of relevance, the color strategy is not effectively able to detect all objects of interest (detects 4), where as when the algorithm uses all strategies, we are able to get significantly better results (detects 7). If we expand the number of proposal boxes from 100 to a larger number (say 1000 or 2000), we are able to get a better performance/recall (0.92/1.0) for both strategies. In general, utilizing selective search algorithm with all strategies would give a better performance than using just the color strategy. Additionally, sometimes the results using color algorithm could be good as well depending upon the image we are trying to run the algorithm on.
 
-**EdgeBoxes:**
+### EdgeBoxes:
 The test image shown above in Appendix B - Image 5 has been annotated with 13 ground truth boxes. On selecting the proposed bounding boxes with the top 100 scores, we found that the number of proposal boxes that have an IoU greater than 0.5 with any of the ground truth boxes is 12. Out of these 12 proposal boxes, only four ground truth boxes have an overlap with atleast one proposal box. The recall for this image using edgeboxes algorithm is 0.307.
 
 Following are a few cases of experimenting with different values of alpha and beta:
@@ -120,7 +122,7 @@ Case 3: Setting alpha and beta as 0.8 had a poorer recall than case 2 , but it g
 
 Setting alpha as 0.5 and beta as 0.5 seemed to give me the best performance in terms of recall.
 
-**Comparison between EdgeBoxes and Selective Search:**
+### Comparison between EdgeBoxes and Selective Search:
 EdgeBoxes for this particular image (Image 5) did not give a significantly improved performance than selective search. However looking at all the images in appendix A and appendix B, edgeboxes in general gave a better performance than selective search.
 I noticed that the proposal boxes generated using edgeboxes were larger in size than the proposal boxes using selective search algorithm because the edgeboxes algorithm tried to detect those boxes which have a lot of contours inside them and scored them highly. So the top 100 scores assigned by the edgeboxes algorithm were generally the larger bounding boxes and thus they had a higher probability of having an IoU > 0.5 with any of the ground truth boxes.