how's the speed compares to squeezeseg?

[lucasjinreal]

How's the speed compares to squeezeseg fo second? I am currently using a cnnseg method got a 80ms speed result:

[traveller59]

in 1080Ti: SECOND 0.04s SECOND v1.5: 0.035s SECOND v1.5 lite: ~0.02s (remove all subm) the time is average inference time over KITTI val set.

[lucasjinreal]

@traveller does second lite also opensourced ?

[ghimiredhikura]

Hello @jinfagang,

Can you post the link of squeezeseg source. Thanks

[traveller59]

Just remove all subm layers... It's very simple. train a lite model with super converge only takes less than 3 hours. note that your picture detect a large area, SECOND will run much slower with more point due to sparse convolution is slow with too much point. When I change the range to [-50, -50, 50, 50] (regular setting in real auto driving system), the time of SECOND v1.5 increase to 66 ms. The SECONV v1.5 lite takes ~30ms because it use less sparse convolution layers. the SECOND v1.5 lite KITTI val AP performance: Car AP@0.70, 0.70, 0.70: bbox AP:90.64, 89.27, 87.88 bev AP:90.18, 87.34, 86.38 3d AP:88.10, 77.68, 75.35 I will upload a checkpoint for lite model, and add raw lidar data support in several days.

[lucasjinreal]

@traveller59 Have u counted in preprocess time ? I see that SECOND have a very long preprocess to gather input data (which can be simpler maybe?). BTW, what 's the unit of [-50, 50]? pixls or real distance in cloud? (My cloud show above only range about 30 circle in real distance in point cloud)

[traveller59]

@jinfagang real distance in cloud in meters. All time measurements are include prep time, measured without async preprocess. The inference preprocess only contains point to voxel operation. other code only used during training.

[lucasjinreal]

@traveller59 That's really promising