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

TL;DR

CaMiT tackles the problem of temporal drift in fine-grained car model recognition by introducing a time-aware dataset with extensive year-level metadata. It evaluates static pretraining, time-incremental pretraining (TIP), time-incremental classifier learning (TICL), and time-aware image generation (TAIG), showing that TICL provides the strongest temporal robustness and TAIG improves generation realism when year information is included. The dataset enables controlled study of how appearances evolve and how models can adapt with minimal retraining, using metrics like $A_{avg}$, $A_{crt}$, $A_{bck}$, and $A_{fwd}$ for classification and $KID$ for generation. By releasing CaMiT and accompanying code, the work provides a practical benchmark to integrate temporal reasoning into fine-grained visual recognition and generation, reducing model obsolescence over long-lasting artifact domains.

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.

Figures (5)

• Figure 1: Illustration of CaMiT content and experiments. (a) exemplifies the Citroën C3 data shift in time. (b) shows that the averaged kernel-inception distance between class embeddings increases with time. (c) highlights the positive effects of time-aware classification over static pretraining.
• Figure 2: Overview of the CaMiT curation pipeline. The pipeline consists of three stages: (1) Data Collection — Flickr queries by car subtype, brand, and model yield 7.5 M candidate images. (2) Data Filtering — automated filtering and de-duplication remove non-relevant or low-quality samples, resulting in 5.87 M car crops. (3) Data Annotation — semi-automatic labeling with Qwen, GPT-4o, and DeiT produces 787 K verified car crops.
• Figure 3: Dataset analysis across time. We compute CLIP ViT-B embeddings and evaluate the kernel-inception distance (KID) between yearly embeddings.
• Figure 4: Static pretraining accuracy (in %) for all train-test year combinations.
• Figure 5: Detailed accuracy (in %) for classification across years using a ViT-S backbone.