Python code to find the roi

import cv2                                                                                                                                                               
import os, sys                                                                                                                                                           
import numpy as np                                                                                                                                                       
v_cap = cv2.VideoCapture(sys.argv[1])                                                                                                                                    
v_len = int(v_cap.get(cv2.CAP_PROP_FRAME_COUNT))                                                                                                                         
for i in range(v_len):                                                                                                                                                   
    ret, frame =                                                                                                                                            
    th = cv2.inRange(frame, (9, 13, 104), (98, 143, 255))                                                                                                                
    points = np.where(th>0)                                                                                                                                              
    p2 = zip(points[0], points[1])                                                                                                                                       
    p2 = [p for p in p2]                                                                                                                                                 
    rect = cv2.boundingRect(np.float32(p2))                                                                                                                              
    cv2.rectangle(th, (rect[1], rect[0]), (rect[1]+rect[3], rect[0]+rect[2]), 255)                                                                                       
    cv2.imshow("t", th)                                                                                                                                                  

I dont think this is quite capturing the region around the segmented area. Don’t we need to do some kind of blob analysis like in the Matlab version?

No need for blob analysis. Actually i later discovered that with changing the 9 to 0 you get very good solid line. You can now use findcontours.

Were you able to process all images using that thresholding function? Are you sure that it crops the required region? It will be helpful if you can please upload the updated code snippet.

Yes, even with this i got perfect cropping. Did you try it? If you find anything not cropped properly, let me know.

I was able to write my own cropping function. Did you ever check the performance of the models without cropping i.e using the full image?

Yes. It didn’t do very well, which makes sense, although ymmv… It depends probably on the model, etc

Hello. I just jumped in the competition. I studied machin learning, but to the industry. So I am trying to gain real world experience here.