Re-produced result is not same as original paper one.

sfzhang15 / RefineDet

Single-Shot Refinement Neural Network for Object Detection, CVPR, 2018

Other

1.44k stars 393 forks source link

Re-produced result is not same as original paper one. #153

Closed youngwanLEE closed 5 years ago

youngwanLEE commented 5 years ago

Hi, I wondered this training script produces the reported result and just tried to train using this script on 4GPU TITAN X server with same setting.

The result is below in which the mAP is lower than 80 mAP that is reported.

How do you think this result? Let me know what I miss.

Thanks

sfzhang15 commented 5 years ago

@youngwanLEE The training has a little fluctuation due to randomness. You can set the IoU threshold as 0.3 (negative) and 0.7 (positive) in ARM to get more stable results.

youngwanLEE commented 5 years ago

@sfzhang15, thanks for reply. Could you notice where I correct exactly?

sfzhang15 commented 5 years ago

@youngwanLEE Adding these two lines after here

multibox_loss_param_arm['overlap_threshold'] = 0.7
multibox_loss_param_arm['neg_overlap'] = 0.3

youngwanLEE commented 5 years ago

@sfzhang15 , Thanks:)