psnr is poor on my custom data

[hawl666]

I train on my own dataset and get a very low psnr.
[图片]

Iter: 99700 PSNR: 22.95 NRays: 5312 OctSamples: 54.7 Samples: 103.3 MeaningfulSamples: 49.2 IPS: 39.1 LR: 0.0010
Iter: 99750 PSNR: 23.06 NRays: 5328 OctSamples: 55.0 Samples: 104.0 MeaningfulSamples: 49.2 IPS: 39.1 LR: 0.0010
Iter: 99800 PSNR: 22.93 NRays: 5312 OctSamples: 54.9 Samples: 103.8 MeaningfulSamples: 49.4 IPS: 38.7 LR: 0.0010
Iter: 99850 PSNR: 22.88 NRays: 5328 OctSamples: 54.8 Samples: 103.4 MeaningfulSamples: 49.1 IPS: 39.2 LR: 0.0010
Iter: 99900 PSNR: 23.02 NRays: 5328 OctSamples: 54.9 Samples: 103.7 MeaningfulSamples: 49.1 IPS: 38.8 LR: 0.0010
Iter: 99950 PSNR: 22.92 NRays: 5312 OctSamples: 54.8 Samples: 103.5 MeaningfulSamples: 49.2 IPS: 39.1 LR: 0.0010
Iter: 100000 PSNR: 22.90 NRays: 5312 OctSamples: 54.7 Samples: 103.4 MeaningfulSamples: 49.3 IPS: 4.8 LR: 0.0010
Train done, test.
0: 12.030964
8: 14.907398
16: 12.652363
24: 13.218551
32: 11.951537
40: 12.5848675
48: 11.751997
56: 14.919983
64: 11.584293
72: 10.137632
80: 14.72267
88: 15.380991
Mean psnr: 12.986938

[hawl666]

result is here.

[Bin-ze]

说一下可能的问题：

1. 图像不匹配，请检查所有的渲染图像，看是否存在gt跟render对不上的情况
2. 自定义的配置问题，请先使用wanjinyou或者wnjinyou_big,进行训练，对比一下是否是这里的问题
3. 你的训练psnr很低，只有22.9，说明两点，一是你的场景很复杂，导致算法学不到好的几何表示，二是姿态不准确，导致重建质量差，请排查一下