Glonce

FWL-ToPM prunes tokens to run full-waveform LiDAR Transformers in real time

researchThe Glonce editorsHugging Face🔥54Glonce rating54Impact40Novelty55Relevance90Surprise25Players15Credibility75
FWL-ToPM prunes tokens to run full-waveform LiDAR Transformers in real time

LiDAR is a key sensor for autonomous driving and robotics, but it struggles in fog and rain, and with transparent objects and low-reflectance surfaces. Full-waveform LiDAR (FWL) is meant to fix this. Instead of a single return, it captures the complete return waveform as a histogram, which preserves detail such as multi-path echoes and weak signals.

Recent Transformer-based approaches have shown promise on FWL data, but the authors name two bottlenecks. The first is high computational cost, caused by redundant tokens in low-intensity background regions. The second is limited generalization, because annotated training data is scarce. The paper tackles both by using the physical structure of FWL data.

The first method is FWL-ToPM, a token pruning and merging mechanism. It uses intensity distributions to selectively cut the number of tokens while keeping the waveform content that matters. The second is FWLAug, a data augmentation framework built as a regularization method. It preserves waveform structure: the Time-of-Flight waveform of each ray and the frustum geometry stay intact. The aim is better generalization to unseen scenes.

On the Ghost-FWL benchmark, the authors report that FWL-ToPM achieves over 8 times speedup at real-time throughput, together with a nearly 9-point F1 gain over the prior state of the art. They call it the first real-time FWL Transformer that also improves accuracy. FWLAug is reported to improve performance further, both on the standard benchmark and on an out-of-distribution test set of unseen scenes. The authors say the two methods push FWL-based perception toward practical deployment. Code and the out-of-distribution dataset are publicly available through the project page at https://keio-csg.github.io/FWL-ToPM.

Key facts

  • FWL-ToPM prunes and merges tokens using intensity distributions, cutting redundant tokens in low-intensity background regions while keeping critical waveform content.
  • On the Ghost-FWL benchmark the authors report an over 8 times speedup at real-time throughput and a nearly 9-point F1 gain over the prior state of the art.
  • FWLAug is a data augmentation framework that keeps each ray's Time-of-Flight waveform and the frustum geometry intact, to generalize better to unseen scenes.
  • FWLAug is reported to improve results on both the standard benchmark and an out-of-distribution test set.
  • Code and the out-of-distribution dataset are public via the project page.

Why it matters

Full-waveform LiDAR keeps far more of the returned signal than conventional LiDAR, which helps in fog, rain and with transparent or dark surfaces. Transformers work well on this data but are expensive, because many tokens come from empty, low-intensity background. The authors say FWL-ToPM is the first real-time FWL Transformer that also gains accuracy, so speed is not bought with a drop in quality.

Who it affects

Researchers working on LiDAR perception and Transformer efficiency are the direct audience. The authors frame the work around autonomous driving and robotics, where LiDAR is a critical sensor and adverse conditions are a known weakness.

How to use it

The code and the out-of-distribution dataset are public through the project page at https://keio-csg.github.io/FWL-ToPM. Teams working on FWL Transformers can try token pruning and merging guided by intensity, and the waveform-preserving augmentation, on their own data.

How solid is it

This is a preprint abstract, and the headline figures come from the authors' own evaluation on the Ghost-FWL benchmark. The abstract does not name the baseline for the 8 times speedup, the hardware, the frame rate that counts as real time, the absolute F1 values, or the prior state-of-the-art method. The size of FWLAug's gain on the out-of-distribution set is also not given. Public code and an open dataset should let others check the claims.

Risks and caveats

The "nearly 9-point" F1 gain is in absolute points, not a relative percentage. The evaluation named is the Ghost-FWL benchmark and an out-of-distribution test set; no on-vehicle or real-world driving deployment is reported. Practical deployment is the authors' aim, not something the abstract demonstrates.