Revisiting the CompCars Dataset for Hierarchical Car Classification: New Annotations, Experiments, and Results.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing car image classification datasets like CompCars have limitations in real-world applicability due to biased data splits.
• Convolutional neural networks (CNNs) achieve high accuracy on fine-grained car classification but performance is not representative of real-world scenarios.
• A need exists for a more robust car classification benchmark that reflects diverse real-world conditions.

Purpose of the Study:

• To create a more realistic and challenging benchmark for car image classification.
• To improve car classification accuracy by utilizing hierarchical annotations and a refined dataset.
• To establish a new baseline for fine-grained car recognition research.

Main Methods:

• Revisiting the CompCars dataset with a new, more representative training/test split.
• Propagating existing type-level annotations and adding car-tight bounding boxes using an automated car detector.
• Designing and implementing three car classification approaches, including two that exploit hierarchical data.

Main Results:

• Established a realistic baseline accuracy of 61% on the revisited CompCars dataset.
• Achieved 70% accuracy in fine-grained car classification by leveraging hierarchical car type information.
• Demonstrated that higher-level classification positively impacts finer-grain classification performance.

Conclusions:

• The revisited CompCars dataset provides a more realistic benchmark for car image classification.
• Exploiting hierarchical car annotations significantly improves classification accuracy.
• The enriched dataset and baseline results pave the way for future advancements in fine-grained visual recognition.