Research map / Detection and segmentation / CNN features
On-Road Object Importance Estimation: A New Dataset and A Model with Multi-Fold Top-Down Guidance
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.
Introduces the TOI on-road object importance dataset (re-annotated KITTI) and a model fusing bottom-up video features with driver-intention, semantic-context, and traffic-rule top-down guidance.
Problem
Estimating the importance of on-road objects from driver-perspective video is under-explored, largely due to the scarcity of public large-scale datasets and the limited capacity of existing bottom-up or single-guidance methods to handle dynamic traffic scenes.
Approach
A model that interactively fuses bottom-up spatio-temporal object features with three top-down guidance factors: driver intention (ego angular-velocity-selected intention masks), semantic context (segmentation-map features), and traffic rule (object-lane cross-attention with an adaptive penalty weighting).
Key results
- On TOI, the proposed model reaches 60 AP and 54 F1 versus 50 AP and 49 F1 for the best baseline Goal.
- On Ohn-Bar, the model reaches 64 AP, a 23.1% improvement over Goal (52 AP).
- Ablation: removing the traffic-rule and driver-intention/semantics modules drops AP from 60 to 20 (bottom-up only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the full multi-fold top-down guidance (TRG and DISG) to bottom-up features (ablation) | On-road importance estimation model · TOI | AP: 20 -> 60 | Helped |
| add TRG to bottom-up features (ablation) | On-road importance estimation model · TOI | AP: 20 -> 31 | Helped |
| add DISG to bottom-up features (ablation) | On-road importance estimation model · TOI | AP: 20 -> 52 | Helped |
| add semantic guiding feature (ablation) | DISG · TOI | AP: 31 -> 49 | Helped |
| add intention guiding mask (ablation) | DISG · TOI | AP: 31 -> 35 | Helped |
| add object-lane interaction without interaction weighting (ablation) | TRG · TOI | AP: 52 -> 46 | Hurt |
| add object-lane interaction weighting to object-lane interaction (ablation) | TRG · TOI | AP: 46 -> 60 | Helped |
| add spatial object features (ablation) | OFE · TOI | AP: 14 -> 60 | Helped |
| add temporal object features (ablation) | OFE · TOI | AP: 34 -> 60 | Helped |
| replace the learnable random-initialized intention mask with the predefined intention mask (ablation) | DISG · TOI | AP | Helped |
| vary weighting parameter α among the three tested values (ablation) | Object-lane interaction weighting · TOI | AP: 57; 55; 60 | No clear effect |
| use the full proposed model instead of Goal | On-road importance estimation · TOI | AP: 50 -> 60 | Helped |
Surprising findings
- Object-lane interaction without the adaptive weighting decreases AP (46 vs 52 without it), suggesting uniform lane interaction hurts performance.
- A learnable intention mask with random initialization did not work, so predefined intention masks were used instead.
Limitations
- Lane marking detection failures degrade performance, e.g., when ego-car moves slowly and lanes are not detected.
- Only three intention behaviors (turning left, going straight, turning right) are modeled.
- Ground-truth object bounding boxes are assumed known, skipping the detection stage.
Experimental setup
Datasets: TOI Ohn-Bar
Benchmarks: TOI Ohn-Bar
Metrics: AP F1 Accuracy
Baselines: Ohn-Bar Goal Zhang Li MENet A2S PGNet Yolo
Builds on: ResNet ROI pooling LSTM CLRNet DeepLabv3
Closest papers by meaning
- Leveraging Driver Field-of-View for Multimodal Ego-Trajectory Prediction (ICLR 2025 · similarity 0.85)
- TUMTraf VideoQA: Dataset and Benchmark for Unified Spatio-Temporal Video Understanding in Traffic Scenes (ICML 2025 · similarity 0.85)
- Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection (AAAI 2024 · similarity 0.84)
- Traffic Scene Parsing through the TSP6K Dataset (CVPR 2024 · similarity 0.84)
- DriveGazen: Event-Based Driving Status Recognition Using Conventional Camera (AAAI 2025 · similarity 0.84)
- Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map (CVPR 2025 · similarity 0.84)
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.