Research map / Detection and segmentation / 3D detection and driving
MTA: Multimodal Task Alignment for BEV Perception and Captioning
Yunsheng Ma, Burhaneddin Yaman, Xin Ye, Jingru Luo, Feng Tao, Abhirup Mallik, Ziran Wang, Liu Ren
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add BLA and DCA alignment losses (MTA) to TOD3Cap | TOD3Cap (Llama 3.2) · TOD3Cap captioning m@IoU=0.5 | CIDEr: 108.7 -> 118.7 | Improved |
| add BLA and DCA alignment losses (MTA) to TOD3Cap | TOD3Cap (Llama 3.2) · nuScenes val detection | mAP: 26.6 -> 27.9 | Improved |
| add BLA and DCA alignment losses (MTA) to TOD3Cap | TOD3Cap (Llama 3.2) · nuScenes val rare classes | mAP: 17.1 -> 18.9 | Improved |
| add BLA and DCA alignment losses (MTA) to TOD3Cap | TOD3Cap (InternLM2) · TOD3Cap captioning m@IoU=0.5 | CIDEr: 105.1 -> 113.0 | Improved |
| add multi-task captioning (vanilla TOD3Cap) to BEVFormer | BEVFormer · nuScenes val detection | mAP: 26.8 -> 26.6 | No clear effect |
| add BLA module only (ablation) | TOD3Cap (Llama 3.2) · TOD3Cap and nuScenes | NDS / CIDEr@0.5: NDS 37.7 -> 38.9; C@0.5 108.7 -> 111.9 | Improved |
| add BLA module only (ablation) | TOD3Cap (Llama 3.2) · TOD3Cap | BLEU-4@0.5: 46.7 -> 46.4 | No clear effect |
| add DCA module only (ablation) | TOD3Cap (Llama 3.2) · TOD3Cap | CIDEr@0.5: 108.7 -> 113.6 | Improved |
| attach BLA at middle Q-Former layer instead of first or last (ablation) | BLA · TOD3Cap | CIDEr@0.5: 111.2 (first), 107.7 (last) -> 111.9 (middle) | Improved |
| use MSE rather than CLIP or cosine similarity for BLA (ablation) | BLA · TOD3Cap | CIDEr@0.5: 110.6 (CLIP), 110.3 (Cos) -> 111.9 (MSE) | Improved |
| use CLIP contrastive loss rather than MSE or cosine similarity for DCA (ablation) | DCA · TOD3Cap | CIDEr@0.5: 112.4 (MSE), 113.1 (Cos) -> 113.6 (CLIP) | Improved |
| add both BLA and DCA instead of one (ablation) | MTA · TOD3Cap and nuScenes | CIDEr@0.5: 111.9 / 113.6 -> 118.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Vote2Cap-DETR — Vote2Cap-DETR++: Decoupling Localization and Describing for End-to-End 3D Dense Captioning (TPAMI 2024) · better on 2 of 2 shared results
- BEVFormer — BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers (ECCV 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment (CVPR 2025 · similarity 0.88)
- MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI (ICML 2024 · similarity 0.88)
- Seeing is Understanding: Unlocking Causal Attention into Modality-Mutual Attention for Multimodal LLMs (ICML 2026 · similarity 0.87)
- Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models (NeurIPS 2025 · similarity 0.87)
- Same or Not? Enhancing Visual Perception in Vision-Language Models (CVPR 2026 · similarity 0.87)
- Vision-Language Models Create Cross-Modal Task Representations (ICML 2025 · similarity 0.87)
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.
