mipnerf360 garden result is poor

[tianxiaguixin002]

I trained the code on mipnerf360 garden data whth following scripts:

python scripts/llff2poses.py --data_dir=./data/mip-360/garden

python scripts/run.py --config-name=nerf-360 \ dataset_name=./data/mip-360 \ case_name=garden mode=train

I have modify the factor=4 in nerf-360.yaml, when finish training, the log show: ..... n_nodes_compacted is value 735094 treenodes.size() is value 735094 Iter: 19050 PSNR: 20.47 NRays: 8080 OctSamples: 46.0 Samples: 119.9 MeaningfulSamples: 32.4 IPS: 23.2 LR: 0.0011 Iter: 19100 PSNR: 20.44 NRays: 8048 OctSamples: 45.9 Samples: 119.7 MeaningfulSamples: 32.5 IPS: 22.6 LR: 0.0010 Iter: 19150 PSNR: 20.41 NRays: 8080 OctSamples: 45.9 Samples: 119.6 MeaningfulSamples: 32.4 IPS: 23.2 LR: 0.0010 Iter: 19200 PSNR: 20.42 NRays: 8080 OctSamples: 46.0 Samples: 119.6 MeaningfulSamples: 32.4 IPS: 22.7 LR: 0.0010 Iter: 19250 PSNR: 20.41 NRays: 8096 OctSamples: 45.9 Samples: 119.5 MeaningfulSamples: 32.3 IPS: 22.8 LR: 0.0010 Iter: 19300 PSNR: 20.44 NRays: 8096 OctSamples: 45.9 Samples: 119.3 MeaningfulSamples: 32.4 IPS: 22.9 LR: 0.0010 Iter: 19350 PSNR: 20.45 NRays: 8128 OctSamples: 45.8 Samples: 119.2 MeaningfulSamples: 32.3 IPS: 22.9 LR: 0.0010 Iter: 19400 PSNR: 20.48 NRays: 8048 OctSamples: 45.7 Samples: 119.1 MeaningfulSamples: 32.5 IPS: 22.7 LR: 0.0010 Iter: 19450 PSNR: 20.46 NRays: 8080 OctSamples: 45.7 Samples: 119.2 MeaningfulSamples: 32.4 IPS: 22.9 LR: 0.0010 Iter: 19500 PSNR: 20.44 NRays: 8096 OctSamples: 45.7 Samples: 119.1 MeaningfulSamples: 32.4 IPS: 22.8 LR: 0.0010 Iter: 19550 PSNR: 20.48 NRays: 8128 OctSamples: 45.7 Samples: 119.0 MeaningfulSamples: 32.2 IPS: 23.1 LR: 0.0010 Iter: 19600 PSNR: 20.47 NRays: 8128 OctSamples: 45.6 Samples: 118.9 MeaningfulSamples: 32.2 IPS: 22.9 LR: 0.0010 Iter: 19650 PSNR: 20.46 NRays: 8144 OctSamples: 45.5 Samples: 118.8 MeaningfulSamples: 32.2 IPS: 22.8 LR: 0.0010 Iter: 19700 PSNR: 20.51 NRays: 8144 OctSamples: 45.6 Samples: 119.0 MeaningfulSamples: 32.1 IPS: 23.1 LR: 0.0010 Iter: 19750 PSNR: 20.46 NRays: 8144 OctSamples: 45.6 Samples: 118.9 MeaningfulSamples: 32.2 IPS: 22.8 LR: 0.0010 Iter: 19800 PSNR: 20.46 NRays: 8128 OctSamples: 45.5 Samples: 118.8 MeaningfulSamples: 32.3 IPS: 23.1 LR: 0.0010 Iter: 19850 PSNR: 20.45 NRays: 8160 OctSamples: 45.5 Samples: 118.8 MeaningfulSamples: 32.2 IPS: 23.3 LR: 0.0010 Iter: 19900 PSNR: 20.48 NRays: 8160 OctSamples: 45.5 Samples: 118.5 MeaningfulSamples: 32.1 IPS: 23.1 LR: 0.0010 Iter: 19950 PSNR: 20.54 NRays: 8176 OctSamples: 45.4 Samples: 118.5 MeaningfulSamples: 32.1 IPS: 23.3 LR: 0.0010 Iter: 20000 PSNR: 20.48 NRays: 8160 OctSamples: 45.4 Samples: 118.5 MeaningfulSamples: 32.1 IPS: 0.8 LR: 0.0010 Train done, test. 0: 10.378842 8: 10.334309 16: 12.714293 24: 8.00071 32: 7.812619 40: 9.196725 48: 12.872265 56: 11.860228 64: 12.492895 72: 13.710265 80: 11.093293 88: 11.43869 96: 14.673181 104: 16.585175 112: 14.040527 120: 10.907651 128: 12.12698 136: 15.01075 144: 12.956245 152: 12.072407 160: 14.45992 168: 9.386232 176: 11.397999 184: 12.814769 Mean psnr: 12.014041

The test psnr is 12? it is a bad result, where is the problem ? thank you.

[Bin-ze]

I think there should be a misalignment between the image and the pose. You can check whether the image_list.txt in the dataset directory is misaligned with the image, or check whether the rendered image does not match the GT. I hope it can help you !

[tianxiaguixin002]

Thank