Research map / Label-efficient and robust learning / Recognition and long-tail
AutoLTS: Automating Cycling Stress Assessment via Contrastive Learning and Spatial Post-processing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AutoLTS, a deep learning framework that predicts cycling level of traffic stress (LTS) from street-view images using an ordinal supervised contrastive loss and a causal-model-based spatial post-processing of road feature predictions.
Problem
Cycling stress assessment (LTS) requires high-resolution road network data (traffic speed, parking, cycling infrastructure) that is slow and costly to collect, hindering large-scale application.
Approach
Combines an ordinal contrastive loss (OrdCon, generalizing supervised contrastive learning with virtual coarser labels) to learn embeddings whose clusters respect the ordering of LTS labels, with a causal-DAG-derived iterative post-processing that enforces spatial smoothness on predicted road features rather than on LTS labels directly.
Key results
- AutoLTS achieves 73.41% LTS accuracy on the random split in data-availability scenario one using only street-view images, versus 70.49% for cross-entropy and 70.75% for SupCon.
- Adding street-view images to partial road features (scenario two) raises accuracy by an average of 23.10 percentage points across splits over CART-based baselines.
- Spatial post-processing improves traffic speed prediction accuracy by 1.01 to 2.05 percentage points, corresponding to 21 to 162 correctly assessed road segments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use street-view images with available road features instead of relying on available road features alone | LTS assessment · LTS scenarios 2 and 3 | Acc: 0.43--32.39 percentage points | Helped |
| Use AutoLTS instead of the strongest baseline in each evaluation setting | AutoLTS · LTS assessment, all scenarios and splits | Acc | Helped |
| Replace the learned LTS prediction module with the exact LTS calculation scheme (ablation) | AutoLTS two-step framework · LTS scenario 1 | Acc: 6.16--40.83 percentage points | Hurt |
| Add spatial post-processing to road feature predictions (ablation) | AutoLTS · Traffic speed prediction, four sub-networks | accuracy: +1.29%; +1.01%; +2.05%; +1.82% | Helped |
| Add spatial post-processing to road feature predictions before exact LTS calculation (ablation) | 2Step-Exact · LTS scenario 1, Random | Acc: 41.97 -> 43.10 | Helped |
| Replace OrdCon with MoCo (ablation) | AutoLTS · LTS scenario 1, Random | Acc: 72.21 -> 73.41 | Hurt |
| Replace OrdCon with SupCon (ablation) | AutoLTS · LTS scenario 1, Random | Acc: 73.30 -> 73.41 | Hurt |
| Add OrdCon instead of SupCon (ablation) | Image-only linear classifier · LTS scenario 1, Random | HLA: 93.41 -> 93.96 | Helped |
| Add OrdCon instead of SupCon (ablation) | Image-only linear classifier · LTS scenario 1, Random | AFR: 11.73 -> 9.95 | Helped |
| Add the ordinal contrastive term using weights (0.95, 0.05), rather than (1.00, 0.00) (ablation) | OrdCon-NN · LTS scenario 1, Random | Acc: 70.75 -> 71.11 | Helped |
| Use OrdCon weights (0.90, 0.10) instead of (0.95, 0.05) (ablation) | OrdCon-NN · LTS scenario 1, Random | Acc: 71.11 -> 69.92 | Hurt |
| Use OrdCon weights (0.85, 0.15) instead of (0.95, 0.05) (ablation) | OrdCon-NN · LTS scenario 1, Random | Acc: 71.11 -> 70.78 | Hurt |
Limitations
- Performance degrades under spatial splits where training and test label distributions differ substantially.
- Experiments use only a single city (Toronto) dataset.
- The model does not account for downstream application-specific error costs.
Experimental setup
Datasets: Toronto street-view images (Google StreetView API) Toronto centerline road network Open Data Canada Open Data Toronto
Benchmarks: Toronto LTS dataset (random, York, Etobicoke, Scarborough splits)
Metrics: Accuracy High/Low-Stress Accuracy (HLA) Average False High/Low-Stress Rate (AFR)
Baselines: Cross-Entropy (ResNet-50) MoCo SupCon CART 2Step-Exact
Builds on: MoCo SupCon CART
Closest papers by meaning
- Road Network Representation Learning with the Third Law of Geography (NeurIPS 2024 · similarity 0.83)
- Being More Lightweight and Practical: Mini-sized Contrastive Learning Pre-trained Models for Fine-grained Traffic Task (ICML 2026 · similarity 0.82)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.81)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.80)
- GEONet: Global Enhancement and Optimization Network for Lane Detection (AAAI 2025 · similarity 0.80)
- I Open at the Close: A Deep Reinforcement Learning Evaluation of Open Streets Initiatives (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.