The result is not match with your given npz file

[lmolhw5252]

Hi,I use your given pth file and your yaml I use test_depth.py file, ckpt/s2d3d_depth_HOHO_depth_dct_efficienthc_TransEn1/ep60.pth config/s2d3d_depth/HOHO_depth_dct_efficienthc_TransEn1.yaml {'mre': array(0.10142188), 'mae': array(0.2026864), 'rmse': array(0.38335027), 'rmse_log': array(0.06684125), 'log10': array(0.04376619), 'delta_1': array(0.90537266), 'delta_2': array(0.96934565), 'delta_3': array(0.98862388)}

mre 163.6033 mae 1.8321 rmse 2.0242 rmse_log 2.1812 log10 2.1632 delta_1 0.0001 delta_2 0.0002 delta_3 0.0005

[sunset1995]

It seem that the data is different from ours. Let use camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_ in area_5a to check the consistency. For the target depth map:

from imageio import imread
depth = imread('data/stanford2D3D/area_5a/depth/camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_depth.png')
print(depth.min(), d.max(), d[d!=65535].max())

It output 356 65535 3156

For the input rgb:

from imageio import imread
rgb = imread('data/stanford2D3D/area_5a/rgb/camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_rgb.png')
print((rgb==0).all(-1).mean())

It output 0.2767658233642578.

[lmolhw5252]

Hi thanks for reply, there is my result:

>>> from imageio import imread
>>> depth = imread('/data1/lhw/workspace/HoHoNet/data/stanford2D3D/area_5a/depth/camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_depth.png')
>>> print(depth.min(), depth.max(), depth[depth!=65535].max())
0 65535 65534
>>> rgb = imread('/data1/lhw/workspace/HoHoNet/data/stanford2D3D/area_5a/rgb/camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_rgb.png')
>>> print((rgb==0).all(-1).mean())
0.0

Different from your result, can you upload your datasets just for depth train and valid? Thanks very much! Have a good day.

[sunset1995]

I think I don't have the right to distribute Stanford2d3d dataset. I download the data from the official site here (http://buildingparser.stanford.edu/dataset.html).

The official RGBs show black in the north-/south-pole region. But from your

>>> print((rgb==0).all(-1).mean())
0.0

It seem that you use the restiched one?

[lmolhw5252]

您好，非常感谢您的解答，但是我还是有一些问题我按照您的https://github.com/sunset1995/HoHoNet/blob/master/README_prepare_data_s2d3d.md 准备的方法，生成对应的文件。

from imageio import imread

rgb = imread('/home/lmolhw/workspace/HoHoNet/Data_test/Stanford/area_5a/rgb/camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_rgb.png')
print((rgb==0).all(-1).mean())

depth = imread('/home/lmolhw/workspace/HoHoNet/Data_test/Stanford/area_5a/depth/camera_e7c078fec2414967b99f65c847b9d38a_office_11_frame_equirectangular_domain_depth.png')
print(depth.min(), depth.max(), depth[depth!=65535].max())

0.25
0 65535 65534

另外还有几个问题：

• 这个数据的prepare对模型的训练结果会有影响么？
• 每个模型大概训练了多少个epoch呢？
• 在test时，(BiFuse's stitching and setting)是什么？
• 训练集和测试集是如何选择的呢？我看原文里面写的是：5用作测试，其余用作训练？
• 或者如果方便，可否提供百度网盘关于depth的训练、测试数据下载～我将非常感激！ 期待您的回复～祝好～