Definition of Block Size and Grid Size during data preparation of Semantic3D

[aldinorizaldy]

Does anyone can explain the definition of block size and grid size in prepare_semantic3d_data.py line 24-25

    parser.add_argument('--block_size', '-b', help='Block size', type=float, default=5.0)
    parser.add_argument('--grid_size', '-g', help='Grid size', type=float, default=0.1)

What is the effect when we use bigger or smaller size to the accuracy?
I'd really appreciate any help. Thank you.

[sldeitz]

@burui11087 I am also interested in this question. I am training with sparse LiDAR point cloud data and am curious about the impact of "block_size" and "max_point_num" in the H5 files. What are the impacts of tuning the parameters of the H5 file?

[aldinorizaldy]

Well, I guess in the case of airborne lidar you need to have a large block size. Otherwise misclassification (building roof labeled as ground) will occur on large building as you can see in my result. I changed the block size into 10 meter, and those errors are 10x10 meter. In that case I would change the block size into something larger.

Max_point_number is, I guess, the sample size in the H5 file. Point cloud is irregular data, in contrast to image, and we need a fix sample size to be trained in the network.

[sldeitz]

@aldinorizaldy thank you! I am testing the classification of curb ramps which have a smaller area. Would it make sense to reduce my block size down to 5 meters?

[aldinorizaldy]

Yes, it does make sense. Actually, the default block size in semantic3d code is 5 m. I think you may start with semantic3d code but use your own data, as semantic3d is outdoor lidar scene.