davideverona / deep-crowd-counting_crowdnet Goto Github PK

Here is my test for density map creation, I draw 10 random points and than create density map, however sum over map is not 10, but 6.95563, for 100 points it's 91.6066 and it depens on point positions, how can we approximate it more precise?

import numpy as np
import scipy
import scipy.spatial
import scipy.ndimage
import cv2
from random import randint

def gaussian_filter_density(mask):
    
    density = np.zeros(mask.shape, dtype=np.float32)
    
    gt_count = np.count_nonzero(mask)
    if gt_count == 0:
        return density

    pts = np.array(zip(np.nonzero(mask)[1], np.nonzero(mask)[0]))
    
    leafsize = 2048
    # build kdtree
    tree = scipy.spatial.KDTree(pts.copy(), leafsize=leafsize)
    # query kdtree
    distances, locations = tree.query(pts, k=2, eps=10.)

    for i, pt in enumerate(pts):
        pt2d = np.zeros(mask.shape, dtype=np.float32)
        pt2d[pt[1],pt[0]] = 1.
        if gt_count > 1:
            sigma = distances[i][1]
        else:
            sigma = np.average(np.array(mask.shape))/2./2. #case: 1 point
            
        density += scipy.ndimage.filters.gaussian_filter(pt2d, sigma, mode='constant')
            
    print "np.sum(density): ", np.sum(density)
    
    return density
    
def test_density():
    h= 100
    w= 120
    mask= np.zeros((h,w))
  
    for i in range(0,10):
        y= randint(0,h-1)
        x= randint(0,w-1)
        mask[y,x]= 1
        
    density= gaussian_filter_density(mask)
    
    cv2.imwrite('mask.png', 255*mask)
    cv2.imwrite('density.png', 255*density)
   
test_density()

HI~
When I calculated the number given by the density map, I found there was a difference between the gt density map before and after downsampling(or resize). So if it is correct to train the model with the downsampled gt density map? Which gt number should be used when evaluating the method? The original number given by the dataset or the number calculated by the downsampled gt density map?

How did you deal with the problem? Thank you.

Hi, I ran the code through and the best iteration was the model saved at 1000 (dcc_crowdnet_train_iter_1000.caffemodel)...
Ground truth Avg: 1642.42 MAE: 812.687
I see David achieved similar error in the pynb on the repo, although a little lower (~700), which is probably due to different test data selected in the 20% test. The paper reported error of ~ 420 over 5 folds. Any idea what the gap may be ? Assuming there is a gap...

Hi @davideverona ,
Thanks for sharing your method.
But i don't know how to make the file of "train.txt", could you please tell me?
It is the best for me if you can provide the "train.txt" and train-datasets.
My email address is [email protected].
Thank you very much!

samylee

Hi @davideverona ,
Could you help me? Please^_^
I realized that the last layer of the output is 28x28, the value of the feature points are less than 1, and even negative numbers. So how do I get the final density map of the image via image mosaic?
It is the best for me if you can provide the code (python or c++). My email address is [email protected].
By the way, The caffemodel that I trained still has some problems, so could you please provide the caffemodel for me? Thanks! And the caffemodel that you trained could be applied to low density populations?
I'm looking forward to your reply. Thank you very much!

Samylee

Hi @davideverona

Can you please share me your ideas to use this code for crowd counting in CCTV cameras?

Thanks
Guru

Hi Davide Verona,

What is the license of this source code?

        for out in net.blobs[out_layer].data:
            y = out[0] #single channel
            Y_deploy.append(y)

the Y_deploy need to be allocated in new memory, or it will be always the last predict result .

            Y_deploy.append(y.copy()) is right.