Research map / Label-efficient and robust learning / Self-supervised and contrastive
CaMiT: A Time-Aware Car Model Dataset for Classification and Generation
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 CaMiT, a Flickr-sourced, time-stamped fine-grained car model dataset (190 models, 787K labeled, 5.1M unlabeled) and evaluates static, time-incremental pretraining, classifier learning, and time-aware image generation.
Problem
Visual models ignore appearance shifts of fine-grained technological artifacts over time; existing temporal datasets are small, coarse-grained, or lack long temporal coverage with fine-grained labels.
Approach
A large, year-stamped fine-grained car dataset built from Flickr with a semi-automatic VLM + specialized-classifier labeling pipeline, used to benchmark time-incremental classification (pretraining, LoRA updates, prototype/classifier updates) and time-aware text-conditioned generation.
Key results
- In static pretraining, LoRA-adapted ViT-B CLIP reaches 65.6% A_avg and MoCo v3 with LoRA reaches 66.0%, while un-adapted DINOv2 ViT-B reaches 26.1%.
- Time-incremental classifier learning with RanPAC on ViT-B MoCo v3+LoRA reaches 87.8% A_avg, over 20 points above the best static pretraining result.
- Time-aware generation (SD1.5_TAIG) lowers KID from 6.83 to 4.19 (x1e-4) and raises A_avg^gen from 46.1% to 54.1% compared with standard SD1.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LoRA adaptation using 2007 data (ablation) | MoCo v3 ViT-S · CaMiT | A_avg: 55.9 -> 64.9 | Helped |
| Add LoRA adaptation using 2007 data (ablation) | MoCo v3 ViT-B · CaMiT | A_avg: 62.4 -> 66.0 | Helped |
| Replace static pretraining with yearly reservoir updates (ablation) | MoCo v3 ViT-S · CaMiT | A_avg: 55.9 -> 58.3 | Helped |
| Use yearly LoRA adaptation on current-year data instead of static pretraining (ablation) | MoCo v3 ViT-S · CaMiT | A_avg: 55.9 -> 76.5 | Helped |
| Add reservoir updates and LoRA on all known years (ablation) | MoCo v3 ViT-S with yearly LoRA · CaMiT | A_avg: 76.5 -> 78.5 | Helped |
| Use yearly LoRA adaptation on current-year data (ablation) | CLIP ViT-B · CaMiT | A_avg: 51.5 -> 76.5 | Helped |
| Replace NCM with incremental prototype updates (NCM-TI) (ablation) | MoCo v3+L_i ViT-S · CaMiT | A_avg: 64.9 -> 71.4 | Helped |
| Use FeCAM instead of NCM (ablation) | MoCo v3+L_i ViT-S · CaMiT | A_avg: 64.9 -> 85.6 | Helped |
| Use RanPAC instead of NCM (ablation) | MoCo v3+L_i ViT-S · CaMiT | A_avg: 64.9 -> 86.6 | Helped |
| Use RanDumb instead of NCM (ablation) | MoCo v3 ViT-S · CaMiT | A_avg: 55.9 -> 82.2 | Helped |
| Include year metadata in fine-tuning captions (ablation) | SD1.5 fine-tuning · CaMiT generated images | KID (×10^-4): 4.48 -> 4.19 | Helped |
| Include year metadata in fine-tuning captions (ablation) | SD1.5 fine-tuning · CaMiT generated images | A_avg^gen: 52.1 -> 54.1 | Helped |
Surprising findings
- Time-incremental classifier updates improve backward accuracy beyond current-year accuracy, since some current samples depict older model variants.
- Specialized in-domain pretraining with LoRA matches large generic models on fine-grained car classification, contrary to prior findings on commonsense-level classes.
Limitations
- Flickr-only sourcing creates selection and geographic bias (Europe dominates brands and models).
- Semi-automatic labeling with VLMs may introduce bias, although manual verification reports 99.6% accuracy on a sample.
- Temporal annotations conflate genuine design change with physical aging of photographed cars.
Experimental setup
Datasets: CaMiT labeled subset CaMiT pretraining subset COCO
Benchmarks: CaMiT
Metrics: A_avg A_crt A_bck A_fwd KID A_avg^gen
Baselines: NCM NCM-TI FeCAM RanDumb RanPAC SD1.5 SD1.5_FT
Builds on: MoCo v3 DINOv2 CLIP LoRA RanPAC RanDumb FeCAM Stable Diffusion 1.5
Closest papers by meaning
- Generating Multimodal Driving Scenes via Next-Scene Prediction (CVPR 2025 · similarity 0.84)
- FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving (NeurIPS 2025 · similarity 0.84)
- Data-efficient Large Vision Models through Sequential Autoregression (ICML 2024 · similarity 0.84)
- Structural Information Guided Multimodal Pre-training for Vehicle-Centric Perception (AAAI 2024 · similarity 0.84)
- STSBench: A Spatio-temporal Scenario Benchmark for Multi-modal Large Language Models in Autonomous Driving (NeurIPS 2025 · similarity 0.84)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 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.