feat(BEVFusion-LiDAR): release BEVFusion-LiDAR 2.7.x

[KSeangTan]

Summary

This PR reintroduces the improvement made by @knzo25 in #16, which shows better results in long-range after performing min-max normalization and fourier features in lidar pointclouds.

Main Changes

• Enable the fourier features in configs. [1], [2], [3]
• Update results in the base model. [4]
• Update results in the j6gen2_base model. [5]
• Add results for the jpntaxi_base model. [6]

Performance

Summary

• Across all datasets and distance ranges, version 2.7.x consistently outperforms 2.6.x.

  • The "Bus" Breakthrough: The most dramatic improvement is seen in the long-range detection (90–121m) of buses. In the Base dataset, the mAP for buses jumped from 0.2849 (v2.6.0) to 0.5178 (v2.7.0)—nearly doubling the performance for large vehicles at a distance.
  • General Stability: While cars and pedestrians show steady incremental gains (usually 1–2), the model's overall heading accuracy (mAPH) scales linearly with its detection accuracy (mAP), suggesting the orientation estimation is robust

• 2.7.x represents a major fix for long-range detection of large vehicles (Buses and Trucks) across all datasets.

---

LargeBus

---

J6Gen2

---

JPNTaxi_Gen2

---

J6Gen2_Base

---

Base

---