The mAP of EfficientDet-D0

[urbaneman]

Hello, @toandaominh1997 , thanks for your great work. I saw your update of the EfficientDet-D0's weight and result. In the table of benchmarking, the title is

PASCAL VOC 2007 (Train/Test: 07trainval/07test, scale=600, ROI Align) The mAP is 31.6. Do you mean the EfficientDet-D0's mAP in VOC07test is 31.6? I see the other implementation's mAP is ≈80 in VOC07test. Are there some troubles or the mAP=31.6 is in COCO Dataset.

[b03505036]

i agree with u. There is a big gap between VOC mAP 0.31 and COCO mAP 0.31

[florisdesmedt]

Did anybody try a COCO training? Just to know the difference between performance of this repo and the performance discussed in the paper.

[HannH]

My test shows it is about 0.3 mAP in Voc 2007 with EfficientDet-D2 too , I guess we need a pretrained model for EfficientDet to achieve mAP in paper.

[heartInsert]

@HannH emmmmmmm , why is just 0.3 map , did you train it enough epoch to convergence ?

[HannH]

@HannH emmmmmmm , why is just 0.3 map , did you train it enough epoch to convergence ?

I test it in both trainset and testset, get the following mAP: test dataset aeroplane: 0 bicycle: 0.5 bird: 0 boat: 0 bottle: 0 bus: 0.0 car: 0.8 cat: 0.5 chair: 0.0625 cow: 0.0 diningtable: 0.0 dog: 0.3333333333333333 horse: 1.0 motorbike: 0.0 person: 0.48801169590643273 pottedplant: 1.0 sheep: 0 sofa: 0.0 train: 0.8333333333333333 tvmonitor: 0 total mAP: 0.27585891812865493

train dataset aeroplane: 1.0 bicycle: 0.6666666666666666 bird: 1.0 boat: 1.0 bottle: 0 bus: 0 car: 1.0 cat: 1.0 chair: 1.0 cow: 0 diningtable: 1.0 dog: 1.0 horse: 1.0 motorbike: 0 person: 0.8421052631578947 pottedplant: 1.0 sheep: 0 sofa: 0 train: 1.0 tvmonitor: 0 total mAP: 0.6254385964912281 train_epoch=62 I dont know why there are some zero mAP objects. But it definitely gets overfitting with only EfficientNet for weight initialization.

[wanglaotou]

i trained my own dataset like voc datasets, but after trained 65 epochs its map is 0.0 and no boxes to nms, have you met this problem?

[gmvidooly]

@wanglaotou same thing happening with me, Retinahead is giving very low cls_score.