About model performance

[iamwilliamli]

Thanks for your work. However, I found it hard to achieve mAP 0.59 with
ap_iou_thresh=0.25, batch_size=4, bn_decay_rate=0.5, bn_decay_step=20, checkpoint_path=None, cluster_sampling='vote_fps', dataset='sunrgbd', dump_dir=None, dump_results=False, learning_rate=0.01, log_dir='log_batch_4', lr_decay_rates='0.1,0.1,0.1', lr_decay_steps='120,160,200', max_epoch=250, no_height=False, num_point=40000, num_target=256, overwrite=False, use_color=False, use_sunrgbd_v2=False, vote_factor=1, weight_decay=0

The highest mAP after 310 epoch is 0.51, Could you please provide a trained model? Thanks a lot for your effort!

[haowei2020]

I have seen your further work《Vote-Based 3D Object Detection with Context Modeling and SOB-3DNMS》. In this paper, you propose the SOB-3DNMS, about this method, I could't understand that how to learn the predicted IoU SM, in other words, how to decide which ground truth bounding box is used to compute the IoU with the m-th proposal bounding box? Can you share the code?