Object-Counting

Computer Vision Research Project

Counting datasets

VGG Cells

• Synthetic dataset;
• 200 images (256x256) containing simulated bacterial cells from fluorescence-light microscopy;
• Number of cells per image: 174 ± 64;
• Ground-truth: dot annotations.
  

CARPK

• 1,448 images (720x1280) of cars captured from different parking lots;
• 90,000 cars;
• Number of cars in a single scene: [1, 188];
• Ground-truth: bounding boxes.
  

ShanghaiTech (Part B)

• Crowd counting dataset;
• 716 images (768x1024) taken from busy streets;
• Number of people in an image: [9, 539];
• Ground-truth: dot annotations.
  

Counting approaches

Quantitative results

VGG Cells

Mean Absolute Error (MAE) and standard deviations for cell counting on VGG Cells dataset.

Method	Details	N = 32	N = 64
Lempitsky and Zisserman (2010)[1]	Linear ridge regression	5.9 ± 0.5	N/A
Lempitsky and Zisserman (2010)[1]	Density learning	3.5 ± 0.2	N/A
Fiaschi et al. (2012)[2]	Regression forests	3.2 ± 0.1	N/A
Arteta et al. (2014)[3]	Interactive counting	3.5 ± 0.1	N/A
Xie et al. (2016)[4]	FCRN-A	2.9 ± 0.2	2.9 ± 0.2
Xie et al. (2016)[4]	FCRN-B	3.3 ± 0.2	3.2 ± 0.2
Cohen et al. (2017)[5]	Count-ception	2.4 ± 0.4	2.3 ± 0.4 (N=50)
Our implementation (without data aug)	---	---	---
FCRN-A	FCRN-A, MSE loss, full images	6.0 ± 1.7	5.4 ± 1.7
FCRN-A	FCRN-A, MAE loss, full images	8.3 ± 2.7	6.6 ± 1.8
FCRN-A	FCRN-A, LogCosh loss, full images	8.3 ± 1.8	3.6 ± 0.3
FCRN-A	FCRN-A, MSE loss, patches 4*(128x128)	5.5 ± 0.5	3.9 ± 1.1
FCRN-A	FCRN-A, MAE loss, patches 4*(128x128)	6.2 ± 1.8	7.4 ± 1.0
FCRN-A	FCRN-A, LogCosh loss, patches 4*(128x128)	5.0 ± 1.8	4.0 ± 0.9
U-Net	U-Net, MSE loss, full images	4.0 ± 0.8	3.5 ± 0.4
U-Net	U-Net, MAE loss, full images	6.8 ± 2.0	5.5 ± 1.5
U-Net	U-Net, LogCosh loss, full images	4.9 ± 1.0	3.4 ± 0.3
U-Net	U-Net, MSE loss, patches 4*(128x128)	4.4 ± 0.9	3.2 ± 0.1
U-Net	U-Net, MAE loss, patches 4*(128x128)	6.6 ± 1.4	6.8 ± 0.9
U-Net	U-Net, LogCosh loss, patches 4*(128x128)	5.5 ± 1.5	3.2 ± 0.4
Our implementation (with data aug)	---	---	---
FCRN-A	FCRN-A, MSE loss, full images	3.2 ± 0.3	2.8 ± 0.3
FCRN-A	FCRN-A, LogCosh loss, full images	3.1 ± 0.3	3.1 ± 0.3
FCRN-A	FCRN-A, MSE loss, patches 4*(128x128)	2.8 ± 0.3	2.7 ± 0.5
FCRN-A	FCRN-A, LogCosh loss, patches 4*(128x128)	3.0 ± 0.3	2.7 ± 0.5
U-Net	U-Net, MSE loss, full images	3.4 ± 0.2	2.8 ± 0.3
U-Net	U-Net, LogCosh loss, full images	3.2 ± 0.3	2.8 ± 0.1
U-Net	U-Net, MSE loss, patches 4*(128x128)	3.1 ± 0.4	3.1 ± 0.5
U-Net	U-Net, LogCosh loss, patches 4*(128x128)	2.8 ± 0.3	2.8 ± 0.2
SegRegNet, density_map	SegRegNet, LogCosh loss, patches 4*(128x128)	N/A	3.0 ± 0.2
SegRegNet, density_map*(seg_map>1e-3)	SegRegNet, LogCosh loss, patches 4*(128x128)	N/A	2.8 ± 0.4

• N - number of train images;
• Our implementation does not include data preprocessing;
• Standard deviation corresponds to 5 different draws of training and validation sets;
• Results are presented just for the test set;
• Counts per image: 174 ± 64.

CARPK

Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Underestimate (%U), Overestimate (%O) and Difference (%D) on CARPK dataset.
MAE, RMSE, %U, %O for LPN, GAP and GSP are taken from Aich et al. (2018) [8].

Method	Details	MAE	RMSE	%U	%O	%D
Hsieh et al. (2017) [6]	LPN	13.72	21.77	N/A	N/A	N/A
Aich et al. (2018) [7]	GAP-224, patches	7.65	9.59	6.56%	0.84%	7.40%
Aich et al. (2018) [7]	GAP-Full, full images	19.61	21.65	18.71%	0.24%	18.95%
Aich et al. (2018) [8]	GSP-224, patches	5.46	8.09	4.14%	1.14%	5.28%
Aich et al. (2018) [8]	GSP-Full, full images	32.94	36.23	31.42%	0.42%	31.84%
Our implementation (without data aug)	---	---	---	---	---	---
FCRN-A	FCRN-A, MSE loss, full images, 15 epochs	21.15	26.34	13.07%	7.38%	20.45%
FCRN-A	FCRN-A, LogCosh loss, full images, 15 epochs	23.34	29.65	20.41%	2.15%	22.56%
FCRN-A	FCRN-A, MSE loss, patches 32 * (128x128), 15 epochs	22.10	28.73	18.13%	3.22%	21.35%
FCRN-A	FCRN-A, MSE loss, patches 32 * (128x128), 3 epochs	14.73	17.45	2.15%	12.08%	14.23%
FCRN-A	FCRN-A, LogCosh loss, patches 32 * (128x128), 15 epochs	26.19	31.63	23.81%	1.50%	25.31%
FCRN-A	FCRN-A, LogCosh loss, patches 32 * (128x128), 5 epochs	12.13	15.72	5.62%	6.10%	11.72%
U-Net	U-Net, MSE loss, full images, 13/15 epochs	17.91	22.75	6.12%	11.19%	17.31%
U-Net	U-Net, LogCosh loss, full images, 12/15 epochs	18.44	23.68	14.28%	3.54%	17.82%
U-Net	U-Net, MSE loss, patches 32 * (128x128), 14/15 epochs	36.29	40.59	34.21%	0.86%	35.07%
U-Net	U-Net, LogCosh loss, patches 32 * (128x128), 15/15 epochs	26.67	30.91	25.18%	0.59%	25.77%
U-Net	U-Net, LogCosh loss, patches 32 * (128x128), 3/3 epochs	14.03	16.49	4.72%	8.83%	13.55%
Our implementation (with data aug)	---	---	---	---	---	---
FCRN-A	FCRN-A, MSE loss, full images, 8/10 epochs	10.41	12.99	8.52%	1.54%	10.06%
FCRN-A	FCRN-A, LogCosh loss, full images, 2/10 epochs	9.95	14.92	7.94%	1.68%	9.62%
FCRN-A	FCRN-A, MSE loss, patches 16 * (256x256), 4/5 epochs	11.18	14.03	8.57%	2.24%	10.81%
FCRN-A	FCRN-A, LogCosh loss, patches 16 * (256x256), 2/5 epochs	12.24	17.02	9.72%	2.10%	11.82%
U-Net	U-Net, MSE loss, full images, 4/10 epochs	10.27	13.36	5.40%	4.53%	9.93%
U-Net	U-Net, LogCosh loss, full images, 1/13 epochs	9.20	11.77	7.11%	1.78%	8.89%
U-Net	U-Net, MSE loss, patches 16 * (256x256), 9/10 epochs	12.77	15.82	11.40%	0.95%	12.35%
U-Net	U-Net, LogCosh loss, patches 16 * (256x256), 2/10 epochs	12.61	16.48	9.88%	2.30%	12.18%
SegRegNet, density_map	SegRegNet, LogCosh loss, full images	8.47	10.71	6.85%	1.34%	8.19%
SegRegNet, density_map*(seg>1e-3)	SegRegNet, LogCosh loss, full images	8.66	10.91	7.10%	1.26%	8.36%

ShanghaiTech (Part B)

Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Underestimate (%U), Overestimate (%O) and Difference (%D) on ShanghaiTech (Part B) dataset.

Method	Details	MAE	RMSE	%U	%O	%D
Zhang et al. (2015) [9]	Crowd CNN	32.0	49.8	N/A	N/A	N/A
Zhang et al. (2016) [10]	MCNN	26.4	41.3	N/A	N/A	N/A
Sam et al. (2017) [11]	Switching CNN	21.6	33.4	N/A	N/A	N/A
Sindagi et al. (2017) [12]	CP-CNN	20.1	30.1	N/A	N/A	N/A
Ranjan et al. (2018) [13]	ic-CNN (one stage)	10.4	16.70	N/A	N/A	N/A
Ranjan et al. (2018) [13]	ic-CNN (two stages)	10.7	16.00	N/A	N/A	N/A
Olmschenk et al. (2019) [14]	MUD-i1NN	13.4	21.4	N/A	N/A	N/A
Mehta and Valloli (2019) [15]	W-Net	6.9	10.3	N/A	N/A	N/A
Our implementation (without data aug)	---	---	---	---	---	---
FCRN-A	FCRN-A, MSE loss, full images, 5/20 epochs	52.95	74.53	19.08%	23.73%	42.81%
FCRN-A	FCRN-A, LogCosh loss, full images, 24/30 epochs	19.95	33.94	11.70%	4.43%	16.13%
FCRN-A	FCRN-A, MSE loss, patches 32 * (128x128), 54/100 epochs	21.49	34.98	9.33%	8.04%	17.37%
FCRN-A	FCRN-A, LogCosh loss, patches 32 * (128x128), 86/100 epochs	20.81	38.11	15.44%	1.38%	16.82%
U-Net	U-Net, MSE loss, full images, 47/50 epochs	24.85	39.06	18.83%	1.26%	20.09%
U-Net	U-Net, LogCosh loss, full images, 31/50 epochs	19.68	35.41	14.00%	1.92%	15.92%
U-Net	U-Net, MSE loss, patches 32 * (128x128), 50/50 epochs	25.02	44.79	16.99%	3.24%	20.23%
U-Net	U-Net, LogCosh loss, patches 32 * (128x128), 50/50 epochs	27.72	45.63	22.08%	0.33%	22.41%
Our implementation (with data aug)	---	---	---	---	---	---
FCRN-A	FCRN-A, MSE loss, full images, 41/50 epochs	20.21	30.97	7.45%	8.89%	16.34%
FCRN-A	FCRN-A, LogCosh loss, full images, 25/50 epochs	19.31	31.82	10.16%	5.45%	15.61%
FCRN-A	FCRN-A, MSE loss, patches 16 * (256x256), 48/50 epochs	19.03	32.80	12.66%	2.72%	15.38%
FCRN-A	FCRN-A, LogCosh loss, patches 16 * (256x256), 48/50 epochs	15.49	27.88	8.87%	3.65%	12.52%
U-Net	U-Net, MSE loss, full images, 36/50 epochs	17.64	28.62	7.92%	6.34%	14.26%
U-Net	U-Net, LogCosh loss, full images, 29/50 epochs	15.88	26.11	8.42%	4.42%	12.84%
U-Net	U-Net, MSE loss, patches 16 * (256x256), 44/50 epochs	18.46	30.63	10.27%	4.65%	14.92%
U-Net	U-Net, LogCosh loss, patches 16 * (256x256), 50/50 epochs	15.20	26.49	7.37%	4.92%	12.29%
SegRegNet, density_map	SegRegNet, LogCosh loss, patches 16 * (256x256)	18.94	30.79	10.10%	5.21%	15.31%
SegRegNet, density_map*(seg>1e-3)	SegRegNet, LogCosh loss, patches 16 * (256x256)	18.83	30.71	9.79%	5.43%	15.22%

Qualitative results (1_FCRN-A, MSE loss)