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

Overview

This repository contains the code and resources for the paper "CaMiT: A Time-Aware Car Model Dataset for Classification and Generation". The dataset and associated code are designed to explore the temporal dynamics of car model representations and evaluate various time-aware learning strategies for fine-grained image classification and generation.

Key Contributions

1. Dataset: Introduces CaMiT, a large-scale, timestamped dataset of car models photographed over 17 years, contributed by 337K unique users.
2. Experiments: Evaluates the effectiveness of static pretraining (SPT), time-incremental pretraining (TIP), time-incremental classifier learning (TICL), and time-aware image generation (TAIG) on the CaMiT dataset.
3. Findings: Demonstrates the importance of time-aware approaches in mitigating temporal data shift and improving performance for both classification and generation tasks.

Repository Structure

The repository is organized into several folders, each corresponding to a specific component of the project. Below is a brief description of each folder:

---

Folder Structure

1. annotation

• Contains the code and scripts used for semi-automatic annotation of the CaMiT dataset.

2. clip_finetune

• Code for finetuning CLIP models with LoRA.

3. incremental

• Contains the implementation of time-incremental classifier learning (TICL) algorithms.

4. mocov3_finetune

• Code for finetuning MoCo v3 models with LoRA for car model classification.

5. preprocessing

• Scripts for detecting the cars from scraped images

6. pretrain

• Code for pretraining MoCoV3 (TIP)

7. time_aware_generation

• Implementation of time-aware image generation (TAIG) using Stable Diffusion.

---

Dataset

• The dataset is available here

Running Experiments

1. Pretrain Mocov3 models with code in pretrain or download checkpoints
2. Navigate to the Incremental folder to run ncm evaluation scripts

Evaluation

• Results and metrics are reported in the paper. To reproduce the results, follow the instructions in the respective folder READMEs.