Faster RCNN training result problem

[hktxt]

I tried to train Faster-RCNN from scratch. The train.py gives me this result:

{ "main/rpn_loc_loss": 0.045347146689891815, "main/rpn_cls_loss": 0.17269396781921387, "main/roi_loc_loss": 0.27309778332710266, "main/roi_cls_loss": 0.49079766869544983, "main/loss": 0.9819365739822388, "lr": 0.0001, "validation/main/map": 0.12205098147362133, "validation/main/ap/aeroplane": 0.2037050348560657, "validation/main/ap/bicycle": 0.15355473865782293, "validation/main/ap/bird": 0.1069814741360702, "validation/main/ap/boat": 0.08437066002057338, "validation/main/ap/bottle": 0.0303030303030303, "validation/main/ap/bus": 0.13091526733063397, "validation/main/ap/car": 0.2179439900145522, "validation/main/ap/cat": 0.1788983192876631, "validation/main/ap/chair": 0.027534965034965036, "validation/main/ap/cow": 0.10386202324629096, "validation/main/ap/diningtable": 0.1376211067035571, "validation/main/ap/dog": 0.16679011471811683, "validation/main/ap/horse": 0.18906283713577696, "validation/main/ap/motorbike": 0.22127384038737857, "validation/main/ap/person": 0.20588034722737675, "validation/main/ap/pottedplant": 0.0101010101010101, "validation/main/ap/sheep": 0.006993006993006993, "validation/main/ap/sofa": 0.03462976657260593, "validation/main/ap/train": 0.13475834327905686, "validation/main/ap/tvmonitor": 0.09583975346687211, "epoch": 13, "iteration": 70000, "elapsed_time": 43275.40101284254 }

the mAP is extremely low... 0.12... I changed nothing, only the path of VOC dataset instead of download. I'am sure the test dataset returns use_difficult=True, return_difficult=True. and DetectionVOCEvaluator() uses use_07_metric=True. These are default setting in train.py. I check some issues, some had a lower mAP than reported. However, my result is wrong... I did not use the pretrained model. To confirm evaluate code, I downloaded the pretrained model from: 'https://chainercv-models.preferred.jp/faster_rcnn_vgg16_voc07_trained_2018_06_01.npz' and tested it, had a result of 0.706. So, it seems like training porblem. Is the epoch of 13 not enough? should I train more epoches?

[Hakuyume]

Could you share the version of ChainerCV and the commit id of the training code?

[hktxt]

@Hakuyume The version is v6. I cloned ChainerCV repo yesterday. So it's the latest commit. https://github.com/chainer/chainercv/commit/c22c4512b1d9d3b8918804dafba1e8d260586321