yolov5

Zone Evaluation on YOLOv5

Download YOLOv5 from https://github.com/ultralytics/yolov5 and install the dependencies. As a reminder, we use YOLOv5-v7.0.

Make sure you have compiled our pycocotools.

Download the pretrained weight file, e.g., YOLOv5-s, from the official website.

Put our provided file zone-val.py to ./yourpath/yolov5/.

Then run,

 python zone-val.py --data coco.yaml --img 640  --weight yolov5s.pt --device 0,1

And you will see the zone evaluation results (YOLOv5-s):

Zone:, ZP, ZP50, ZP75, ZPs, ZPm, ZPl, ZR1, ZR10, ZR100, ZRs, ZRm, ZRl
z05:  [       37.4        57.2        40.2        21.1        42.3          49        31.1        51.6        56.6        37.8        62.5        72.2]
---------------------------------------
z01:  [       28.8        44.3        30.4        19.4        36.8        42.7        34.3        49.4        51.3        40.6        60.5        66.3]
z12:  [       34.9        52.7          38        21.6        41.3        46.3        37.7        53.2        55.1        36.8        60.2        70.8]
z23:  [       36.9          55        39.3        20.7        42.7        49.1        37.2        52.1        53.9        36.8        59.7        67.4]
z34:  [       35.1        53.8        37.2        21.8        41.6        45.9        37.7        50.5        52.1        35.6        58.6        63.2]
z45:  [       38.4        57.8        42.3          27        45.1        47.6        45.6        54.9          56        39.2        61.1          69]
---------------------------------------
ZP_variance:  [     10.488      20.739       15.42      6.6149      7.2542      4.5353      14.179      3.8022      3.1174       3.321     0.69311      6.5777]
SP, SP50, SP75, SPs, SPm, SPl, SR1, SR10, SR100, SRs, SRm, SRl
SP: [       33.3        50.5        35.6        20.9        40.1        45.6        36.7        51.3        53.2        38.1        60.1        67.5]

Test resolution: 640.

Detector	SP	$\text{ZP}^{0,5}$	Variance	$\text{ZP}^{0,1}$	$\text{ZP}^{1,2}$	$\text{ZP}^{2,3}$	$\text{ZP}^{3,4}$	$\text{ZP}^{4,5}$	params (M)	FLOPs (B)
YOLOv5-s	33.3	37.4	10.5	28.8	34.9	36.9	35.1	38.4	7.2	16.5
YOLOv8-n	32.3	37.3	19.7	26.4	34.0	37.0	35.8	39.6	3.2	8.7
YOLOv5-m	40.8	45.2	12.9	36.0	42.3	44.5	43.2	46.7	21.2	49.0
YOLOv8-s	39.8	44.9	24.4	33.4	42.2	44.3	43.2	48.5	11.2	28.6

Discussion: If we compare YOLOv8 and YOLOv5 with similar AP, the improvement of YOLOv8 mainly comes from large objects and central zone. Besides, YOLOv5 performs better in spatial equilibrium (lower variance).