search

BichenWuUCB / SqueezeSeg

Implementation of SqueezeSeg, convolutional neural networks for LiDAR point clout segmentation
BSD 2-Clause "Simplified" License
566 stars 239 forks source link

hello，Why can't I get the same results as paper with the lidar-2d dataset，is there something wrong? #46

Open YukioSama opened 4 years ago

YukioSama commented 4 years ago

I get much lower results than paper

zwqnju commented 4 years ago

My result was also not as good as the paper. In the paper, the class-level result is: Class Precision Recall IOU Car 66.7 95.4 64.6 Ped 45.2 29.7 21.8 Cyclist 35.7 45.8 25.1

If using the model file in this, the result is: car: Pixel-seg: P: 0.656, R: 0.965, IoU: 0.641 pedestrian: Pixel-seg: P: 0.328, R: 0.290, IoU: 0.182 cyclist: Pixel-seg: P: 0.314, R: 0.610, IoU: 0.261

However, I trained the model and got the following result on the validation set: car: Pixel-seg: P: 0.666, R: 0.960, IoU: 0.648 pedestrian: Pixel-seg: P: 0.198, R: 0.359, IoU: 0.146 cyclist: Pixel-seg: P: 0.274, R: 0.611, IoU: 0.233