CaMiT: A Time-Aware Car Model Dataset for Classification and Generation

AI summary

Overview

Research area: Computer vision, specifically fine-grained visual categorization, continual learning, and time-aware image generation.

Technical level: Intermediate. The concepts are explained accessibly, but the experiments use continual-learning terminology (backward/forward transfer, class-incremental learning) and parameter-efficient adaptation methods.

Scope: The paper introduces CaMiT, a car-model image dataset with year-level timestamps, and uses it to measure how classification and generation degrade as visual classes change appearance over time, testing three classification settings and one generation setting as mitigations.

What This Paper Is About

Most large visual datasets assume a one-off training paradigm and ignore the fact that visual classes change appearance over time. For technological artifacts such as cars, designs, variants, and camera hardware shift continuously, so a model trained on one year's images becomes less accurate when tested on images from earlier or later years. The paper builds a large, year-annotated car dataset spanning roughly two decades and uses it to quantify that temporal shift and to test strategies that reduce it, in both classification and image generation.

Key Contributions

1. CaMiT, a time-aware fine-grained dataset. A labeled subset of 787K samples covering 190 car models (48 brands) published between 2007 and 2023, plus an unlabeled pretraining subset of 5.1M car crops spanning 2005 to 2023, collected from Flickr and built with a semi-automatic filtering and annotation pipeline.

2. A systematic evaluation of temporal data shift in classification. Three settings are compared: static pretraining (SPT), time-incremental pretraining (TIP), and time-incremental classifier learning (TICL), with accuracy aggregated over all, current, backward, and forward train-test year combinations (A_avg, A_crt, A_bck, A_fwd).

3. A time-aware image generation (TAIG) task. Temporal metadata (the photo publication year) is inserted into training captions, and the effect on generation quality and coherence is measured.

4. Public release of dataset and code, with the dataset distributed as image links, embeddings, and metadata rather than the images themselves.

Main Findings

• Temporal shift is measurable and grows with the year gap. Kernel-inception distance (KID) between yearly class embeddings increases with the time gap, and the static pretraining accuracy gap between train and test years generally grows as the difference between those years grows. The authors illustrate change points such as the Citroën C3 III in 2016 or the second generation of the Volvo XC90 in 2015, and note that the least dynamic models include the Citroën 2CV, Porsche 356, Ferrari 599, Lancia Delta, and Fiat 500.

• Specialized pretraining is competitive with large generalist models. Under static pretraining with ViT-B, MoCo v3 with LoRA adaptation on 2007 data reaches A_avg 66.0, compared with 65.6 for CLIP+LoRA adaptation and 62.4 for unadapted MoCo v3. Unadapted CLIP reaches 51.5 and DINOv2 26.1. With ViT-S, MoCo v3+LoRA reaches 64.9 versus 20.9 for DINOv2. The authors state this differs from earlier findings where generic pretraining was much better for commonsense-level classes.

• Accuracy degrades in both temporal directions without mitigation. A_crt is consistently higher than A_bck and A_fwd; for MoCo v3 (ViT-S) the values are 65.9, 54.0, and 56.6 respectively, and for MoCo v3+LoRA (ViT-B) they are 76.5, 63.2, and 67.4.

• Time-incremental pretraining helps, and LoRA beats reservoir updates. With ViT-S, reservoir-based updates reach A_avg 58.3, LoRA adaptation using only the current year reaches 76.5, and combining reservoir training with LoRA over all known years reaches 78.5. With ViT-B, CLIP+LoRA reaches 76.5 and MoCo v3+LoRA 75.3. Differences versus static pretraining are statistically significant (p < 0.01). LoRA and reservoir yearly updates cost 0.3 and 18 GPU hours respectively.

• Time-incremental classifier learning gives the best accuracy. The A_avg gain over the best static pretraining results exceeds 20 points for both ViT architectures, and improvements over the best TIP results reach 8.1 points for ViT-S and 11.3 points for ViT-B, all statistically significant (p < 0.01). The best overall result is RanPAC on a LoRA-adapted MoCo v3 backbone, at 86.6 (ViT-S) and 87.8 (ViT-B). RanPAC is best in most cases, except for unadapted MoCo v3 where RanDumb is better (82.2 for ViT-S, 81.4 for ViT-B). TICL is especially useful for DINOv2, whose accuracy roughly triples (e.g., 20.9 to 66.4 for ViT-S).

• Positive backward transfer appears in a realistic time-aware setting. Improvements are particularly large for past data, with past scores often surpassing current accuracy; the authors explain this by noting that some current samples depict older model variants. This finding was not previously reported in the continual learning literature. Positive forward transfer also occurs but shrinks as the future gap grows.

• TICL gains come with a memory cost. FeCAM, RanPAC, and RanDumb require over 100M additional parameters; RanPAC and RanDumb need a projection matrix with r = 10,000 while d = 384 or 768. Yearly updates take between 3 and 5 minutes.

• Time-aware captions improve generation. SD1.5_TAIG achieves KID 4.19 and A_avg^gen 54.1, better than SD1.5_FT (KID 4.48, 52.1) and standard SD1.5 (KID 6.83, 46.1). Both improvements over standard generation are statistically significant (p < 0.01), and the year-agnostic fine-tuning gains are smaller than the time-aware ones even though SD1.5_FT aggregates training images across all years.

Methodology in Plain English

The researchers chose car models as a representative technological artifact whose design visibly changes over time, and Flickr as a source because it is the only public source they found that allows collection over nearly 20 years. They issued 425 unique queries with car subtypes, brands, and models, using Flickr's temporal filter, taking up to 5,000 images per query per year from 2005 to 2023, which produced 7.5M candidate images.

Filtering removed duplicates using Euclidean distance between CLIP embeddings with a threshold of 0.9, kept detections from YOLOv11x (pretrained on COCO) with confidence at least 0.6 and bounding boxes of at least 64 pixels per side, used Qwen2.5-7B to remove interiors, toy models, and overly detailed or abstract images, blurred faces, and removed images with strongly overlapping bounding boxes using SAM 2 segmentation. This left 5.87M car crops.

Labeling was semi-automatic. Qwen2.5-7B predicted a model name and confidence against a Wikipedia-derived list, GPT-4o confirmed or proposed an alternative, and two DeiT classifiers (DeiT_Q and DeiT_G) were trained with the VLM predictions as weak labels, five instances each on disjoint 80/20 splits, on top of a MoCo v3 ViT-S encoder pretrained on all filtered crops. Thresholds of 90%, 90%, 80%, and 80% were determined by manually validating 20,000 instances of 20 car models with three annotators (Fleiss' Kappa 0.76), giving 98.8% average accuracy. A final manual verification of 100 random crops (IAA 0.79) led to dropping 14 models that produced five or more errors (69.1% of the error count), leaving 190 models and a final accuracy of 99.6%.

For evaluation, accuracy is aggregated over train year i and test year j as A_avg (all combinations), A_crt (i = j), A_bck (i > j), and A_fwd (i < j). Classification uses ViT-S and ViT-B backbones with DINOv2, CLIP (trained on 2B LAION images), and a car-specific MoCo v3, adapted with LoRA (rank 8, alpha 16) on attention layers. MoCo v3 is pretrained for 200 epochs with AdamW, a cosine schedule, learning rate 1.5e-4, and batch size 2048, then fine-tuned for 60 epochs per year on 200K images with learning rate 1e-4. A nearest-class-mean classifier is trained on each year's labeled data. Generation fine-tunes Stable Diffusion 1.5 with LoRA (rank and alpha 64, batch size 64, 30 epochs, AdamW at 1e-4, FP16, 512x512, checkpoints every 5,000 steps), prompting with captions of the form "A photo of CAR_MODEL in YEAR". Statistical significance is assessed with the Wilcoxon signed-rank test. All experiments used around 1,300 A100 GPU hours.

Why This Matters

The paper shows that treating a visual dataset as fixed and timeless makes models obsolete as visual classes evolve, and it provides a resource where this effect can be measured rather than assumed. It also contradicts a prior result for commonsense-level classes by showing that specialized, in-domain pretraining remains competitive for fine-grained recognition.

Real-world applications:

• Automotive and insurance workflows where vehicle identification must work across model generations and production years.
• Market and trend analysis of vehicle design cycles, using year-level detection of when new variants become visually dominant.
• Continual learning research on realistic streams with emerging, evolving, and disappearing classes rather than randomly shuffled splits.
• Content generation and media production where synthetic vehicle imagery should reflect the correct time period.

Industry relevance: the finding that LoRA-based yearly updates cost 0.3 GPU hours while reservoir-based updates cost 18 GPU hours, and that classifier-only updates take 3 to 5 minutes, gives practical guidance for teams that need to keep deployed recognition systems current without full retraining. The paper also notes the risk that vehicle image generation could be misused, for example in insurance fraud systems.

Future Directions

• Disentangling genuine design evolution from physical aging, since the current temporal annotations mix new model variants with older cars photographed years after release; the authors suggest aligning timestamps with official model release dates or registration metadata.
• Extending coverage to a more distant past, which the authors note is difficult because the amount of digitized photography is insufficient to represent the car models.
• Refining how temporal awareness is injected into generative pipelines beyond simple caption-level metadata.
• Exploring more sophisticated methodological approaches to the SPT, TIP, TICL, and TAIG scenarios, which the authors explicitly invite other teams to develop.

Target Audience

Researchers and engineers working on fine-grained visual categorization, continual and incremental learning, dataset construction, and time-aware generative models. It is also relevant to practitioners who deploy vehicle recognition or vehicle image generation systems and need to understand how quickly such models decay and what mitigation options are cheapest.

Note on reported figures: the paper's main text and abstract state 190 car models, while the Limitations section describes Europe as dominating with "99 out of 199 models." The 190 figure is used throughout the abstract, contributions, and dataset statistics.

Authors’ abstract

AI systems must adapt to evolving visual environments, especially in domains where object appearances change over time. We introduce Car Models in Time (CaMiT), a fine-grained dataset capturing the temporal evolution of car models, a representative class of technological artifacts. CaMiT includes 787K labeled samples of 190 car models (2007-2023) and 5.1M unlabeled samples (2005-2023), supporting both supervised and self-supervised learning. Static pretraining on in-domain data achieves competitive performance with large-scale generalist models while being more resource-efficient, yet accuracy declines when models are tested across years. To address this, we propose a time-incremental classification setting, a realistic continual learning scenario with emerging, evolving, and disappearing classes. We evaluate two strategies: time-incremental pretraining, which updates the backbone, and time-incremental classifier learning, which updates only the final layer, both improving temporal robustness. Finally, we explore time-aware image generation that leverages temporal metadata during training, yielding more realistic outputs. CaMiT offers a rich benchmark for studying temporal adaptation in fine-grained visual recognition and generation.