How to obtain better KITTI results?

[hungdche]

Hi,

Thanks for such an incredible project! I've tried running KITTI odometry sequence 0 and obtain the result as follows:

Although it looks quite nice, it seems like it can achieve better quality (as mentioned in #3) as @YuePanEdward has gotten a much better looking result on the same sequence:

I'm not sure if the issue lies on my parameters, so I will post them down below:

field = reconstructor.reconstruct(
        input_xyz, sensor=input_sensor, detail_level=None,
        # Minor configs for better efficiency (not necessary)
        approx_kernel_grad=True, solver_tol=1e-4, fused_mode=True,
        # Chunked reconstruction (if OOM)
        chunk_size=51.2,
        preprocess_fn=nksr.get_estimate_normal_preprocess_fn(64, 85.0)
    )

I would appreciate any help! Thank you in advance.

[heiwang1997]

Hey I'm not pretty sure what exact parameters are used (perhaps @YuePanEdward could kindly advise :) ) But give your results, what you could do is perhaps to increase the knn of your normal estimation search, i.e using

preprocess_fn=nksr.get_estimate_normal_preprocess_fn(96, 80.0)   # <- 96 could be larger while 80.0 could be smaller

You can also cut off the points that are too far away from the sensor origin, because the LiDAR beam's noise increase as it travels longer distance :)

[hungdche]

Hi @heiwang1997, thanks for the advice! Unfortunately increase the knn does not help much, but I was able to get a (subjectively) better mesh by increasing the voxel size to 0.15 I'll mark this as close.