Table 3.
Summary of 2D-based intra-frame compression methods.
| Category | Method | Type | Main Features | Setup | Dataset | Performance | Source Code |
|---|---|---|---|---|---|---|---|
| Traditional | PNG (1997) [59] | Lossless | Applies PNG to range images from LiDAR point clouds | Not disclosed | Velodyne and Ibeo sensors | bpp: 7.5–15.2 [58] PSNR: 111 [39] | Open source |
| JPEG-LS (2000) [61] | Lossless | Applies JPEG-LS to range images from LiDAR point clouds | Not disclosed | Velodyne and Ibeo sensors | bpp: 6.4–22.4 [58] PSNR: 110 [39] | Open source | |
| CLUSTER (2019) [67] | Lossless/Lossy | Uses the shape of RI’s segmented regions to feed the prediction module | Intel Core i5-6300HQ with 4GB RAM | KITTI | CR: 4.83–30.21 | Not disclosed | |
| SC-CSS (2021) [69] | Lossy | Compresses segments of non-/ground points using a combination of RI and 3D representations | Not disclosed | Velodyne HDL-32E sensor | bpp: 6 | Not disclosed | |
| RAI IC (2022) [65] | Lossless | Uses standard image compression methods on images created from range, azimuth, and intensity | Not disclosed | Velodyne VLP-32C sensor | bpp: 10–17 | Open source | |
| Cylindrical Pred. (2023) [66] | Lossless/Lossy | Deploys a prediction scheme on a Cartesian-to-cylindrical projection for spinning LiDARs | Not disclosed | KITTI | - | Not disclosed | |
| Learning | 2D RNN with RB (2019) [70] | Lossy | Uses a RNN with Residual Blocks on range image-based matrices | Intel Core i7-7820X w/ Nvidia GeForce GTX 1080 | Velodyne HDL-32 sensor | bpp: 2.04–4.046 | Not disclosed |
| HSC (2021) [73] | Lossy | Applies Draco [77] on semantic segments provided by RangeNet++ [72] | Intel Core i7-7700K w/ Nvidia TITAN RTX and 32GB RAM | SemanticKITTI | bpp: 0.2–14 PSNR: 30–70 |
Not disclosed | |
| RIC-Net (2022) [71] | Lossless/Lossy | Applies a three stages end-to-end range image-based entropy network | Intel Core i7 w/ Nvidia GeForce GTX 1080Ti | KITTI, Oxford and Campus16 | bpp: 4.1 | Not disclosed | |
| R-PCC (2022) [75] | Lossless/Lossy | Applies real-time sampling segmentation and point-plane mixing modeling to RI | Not disclosed | KITTI, Oxford and HKUSTCampus | bpp: 1.15–5.67 | Open source | |
| SPR (2022) [74] | Lossy | Encodes labels, predictions, and residuals from RangeNet++ [72] RI segments | Intel Core i7-7700K w/ Nvidia GTX 1080Ti | SemanticKITTI | bpp: 6.3–7 | Not disclosed | |
| SCP (2023) [76] | Lossless/Lossy | Offers a framework to convert raw data to spherical coordinates | 2 AMD EPYC 7742 and 8 Nvidia A100 | Ford and SemanticKITTI | - | Not disclosed |