Different Map on same dataset for testing and training (very low map while testing (0.00688) while training gave 0.79 Map)

[bobbilichandu]

I got 0.798 final MAP while testing. But using test.py , I am getting very low MAP score.

Scanning images: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4370/4370 [00:14<00:00, 308.58it/s] Scanning labels data/crowdhuman-608x608.cache (4369 found, 0 missing, 1 empty, 2 duplicate, for 4370 images): 100%|██████████████| 4370/4370 [00:00<00:00, 16752.60it/s] Class Images Targets P R mAP@.5 mAP@.5:.95: 21%|███████▌ | 58/274 [05:45<26:06, 7.25s/it] Class Images Targets P R mAP@.5 mAP@.5:.95: 100%|███████████████████████████████████| 274/274 [24:30<00:00, 3.69s/it] all 4.37e+03 1.83e+05 0.0379 0.0494 0.00688 0.00133 person 4.37e+03 8.26e+04 0.00798 0.011 0.000274 6.39e-05 head 4.37e+03 1.01e+05 0.0679 0.0878 0.0135 0.00259 Speed: 29.4/1.7/31.1 ms inference/NMS/total per 640x640 image at batch-size 16

command used : "python3 test.py --weights runs/exp4_yolov4-p6/weights/best_yolov4-p6.pt --iou-thres 0.50 --data data/coco.yaml --batch-size 16 --save-json --task test --device 1,2 --verbose --save-txt"

What might be the issue?

[WongKinYiu]

what is your training size?

[bobbilichandu]

15000 images, crowdhuman training dataset

[WongKinYiu]

oh, i mean input resolution.

[bobbilichandu]

608*608

[WongKinYiu]

try python3 test.py --weights runs/exp4_yolov4-p6/weights/best_yolov4-p6.pt --iou-thres 0.65 --thresh 0.001 --data data/coco.yaml --batch-size 8 --save-json --task test --device 1 --verbose --save-txt

[Goru1890]

608*608

Is it possible to train the network with images with that size? How does using an input image size smaller than the network size affect the network?