Two questions about inferring the lightstereo model

[Yihan-zz]

Hi, I have two questions while running inference with the lightstereomodel.

1. Is Image Normalization missing in the config file?

It seems the current configuration file for KITTI 2015 evaluation is missing the image normalization step:
{ NAME: NormalizeImage, MEAN: [ 0.485, 0.456, 0.406 ], STD: [ 0.229, 0.224, 0.225 ] }
Reference:

OpenStereo/cfgs/kitti15_eval.yaml

Lines 12 to 14 in 5f0134c

	- { NAME: RightTopPad, SIZE: [ 384, 1248 ]}
	- { NAME: TransposeImage }
	- { NAME: ToTensor }

Without normalization, the results are poor:

After adding normalization, the results significantly improve and match the LightStereo-S (Ours) metrics reported in Table VII of the paper:

2. Performance discrepancy:

Result of LightStereo-S-KITTI.ckpt is better than the paper.

My inference command:

My inference result:

The output shows D1_all = 1.3604, while the paper reports 2.30 for the same metric:

Could you please clarify if I have misinterpreted the evaluation setup or if there is a specific reason for this performance gap?

[ZhangZZ6666]

评估使用的是训练集，已经训练过的权重指标肯定很高。测试集的标签官方并不公开。