I am training a YOLOv4 model to detect people using VisDrone2019 dataset. Training images are large (up to 2000 x 1920) and people relatively small (depends on the image, some images are taken closer than other which leads to larger bounding boxes). My issue is that I keep getting:
class_id = 0, name = person, ap = 0.00% (TP = 0, FP = 14)
for conf_thresh = 0.25, precision = 0.00, recall = 0.00, F1-score = -nan
for conf_thresh = 0.25, TP = 0, FP = 14, FN = 15222, average IoU = 0.00 %
IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
mean average precision (mAP@0.50) = 0.000000, or 0.00 %
I tried calculating custom anchors (6,16, 9,30, 17,26, 13,47, 22,44, 20,73, 37,58, 32,105, 61,151), but, after 2000 iterations, I get loss_loss = -nan
I am training a YOLOv4 model to detect people using VisDrone2019 dataset. Training images are large (up to 2000 x 1920) and people relatively small (depends on the image, some images are taken closer than other which leads to larger bounding boxes). My issue is that I keep getting:
I tried calculating custom anchors
(6,16, 9,30, 17,26, 13,47, 22,44, 20,73, 37,58, 32,105, 61,151)
, but, after 2000 iterations, I getloss_loss = -nan
Here's my configuration file:
PS : I cropped my images into 4 pieces to fit a small sized network and I have patched ignored regions with a white rectangle.
Any help would be appreciated. Thanks !