An Efficient Multi-Scale Transformer Network with Fusion-Attention for Point Cloud-Semantic Segmentation in Urban Environments
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
This article investigates point-cloud segmentation, which is crucial but challenging for scene interpretation, especially for three-dimensional (3D) urban scenes at a city scale. Compared with the previous approaches, the proposed method gains a competitive advantage by leveraging an efficient multi-scale transformer, which complements the convolution in a hierarchical network to improve the representation ability with globally contextual information. More specifically, to address the problem of quadratic complexity that hinders large-scale point-cloud processing, a lightweight attention module with linear complexity is introduced by sequentially implementing channel and spatial attention to replace quadratic dot-product attention. Based on this lightweight attention module, an encoder based on a transformer is implemented to aggregate the feature sequence within a scale into a learnable token. To improve the efficiency of integrating information of multiple scales with no inductive bias, fusion attention is proposed, using only learned tokens to calculate the query, in which the complexity of the attention map can be bounded to be linear. The fusion-attention module is embedded in the multi-scale transformer to further expand the receptive field. The proposed method extends the previous hierarchical networks of point-cloud processing by incorporating the detailed information extracted via convolution and the globally contextual information extracted by the multiscale transformer to greatly improve the representative ability of features for the accurate segmentation of point-cloud data. Two benchmark datasets (Dayton Annotated LiDAR Earth Scan [DALES] and Toronto3D) were used to assess the proposed method. This method achieved an improvement of approximately 1.5% in mean intersection over union for semantic segmentation on the DALES dataset compared with the state-of-the-art methods. Meanwhile, an ablation study showed that consistent improvements were mainly attributed to the wide applicability of the efficient attention mechanism for enlarging the receptive field.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it