Object Detection and Classification of Cars by Make using MMDetection

Summary

This presentation discusses the development of a model for detecting and classifying car makes using MM detection and ResNet. The team worked with the Stanford Cars dataset, faced challenges in real-world generalization due to dataset limitations, and tested various approaches such as fine-tuning pre-trained models and stacking classifiers. While the model showed promising results in controlled environments, it struggled in real-world scenarios. Key insights included the importance of quality data, transfer learning, and the need for more diverse datasets to improve performance. Future improvements aim to enhance model efficiency and generalization for real-time detection.

Takeaways

• 😀 The project aimed to develop a model for detecting and classifying car makes using real-world images.
• 😀 The Stanford Cars dataset was used for training, containing 16,000 images of 196 car makes, models, and years up to 2013.
• 😀 The main challenge was to detect car makes and models accurately in real-world images, given dataset limitations.
• 😀 Initial attempts with pre-trained object detection models (e.g., Faster R-CNN, YOLOX) didn’t yield satisfactory results due to dataset differences.
• 😀 To improve performance, the team stacked a ResNet model on top of existing object detection models, resulting in more accurate bounding boxes.
• 😀 The stacked model achieved over 70% validation accuracy on the Stanford Cars dataset, improving over previous approaches.
• 😀 The data set was limited by outdated images, making it difficult to generalize the trained model to newer cars or uncontrolled environments.
• 😀 The stacked ResNet model showed better bounding box accuracy, but sometimes misclassified car makes in real-world images.
• 😀 Real-world tests, such as on Wilshire Boulevard, showed poor generalization, highlighting the challenge of working with uncontrolled street data.
• 😀 Key lessons learned include the importance of good data quality, the usefulness of transfer learning, and the need for more comprehensive, up-to-date datasets for better performance in real-world applications.

Q & A

What was the primary goal of the project presented in the script?

The primary goal was to develop a model capable of detecting and classifying the make of cars in real-world images using the Stanford Cars dataset.

Which dataset was used for training the model, and what are its key characteristics?

The model was trained using the Stanford Cars dataset, which contains around 16,000 images of 196 different car makes, models, and years, up to 2013.

What were the initial challenges faced when using pre-trained object detection models like Faster R-CNN, Mask R-CNN, and YOLOX?

The initial challenges were that the pre-trained models, which were trained on the Coco dataset, didn’t perform well due to the significant differences between the Coco dataset and the Stanford Cars dataset.

Why did the team decide to limit the class labels to just the make of the car?

The team limited the class labels to only the make of the car (instead of make, model, and year) to simplify the classification task and improve performance due to the large class distribution and nature of the data.

What was the outcome of stacking a ResNet model on top of pre-trained object detectors?

Stacking the ResNet model on top of the pre-trained detectors resulted in better performance, including more accurate bounding boxes and a significant improvement in classification accuracy compared to the fine-tuned models.

What were some of the issues with the Stanford Cars dataset?

The Stanford Cars dataset had several issues, such as being outdated (covering cars only up to 2013), lacking a comprehensive range of brands, and mainly containing images from controlled environments like dealerships, making it challenging to generalize to real-world scenarios.

How did the model perform when applied to real-world images, such as those taken on Wilshire Boulevard?

The model performed poorly on real-world images, like those taken on Wilshire Boulevard, due to the mismatch between the controlled environment in the dataset and the variability found in real-world settings.

What did the team learn about the importance of data for machine learning models from their experience?

The team learned that having the right and diverse data is crucial for effective model performance. While the model performed well on test data, it struggled with real-world images because the training data did not represent the variety of real-world scenarios.

What was the role of transfer learning in the project's success?

Transfer learning was highly beneficial, as pre-trained models (such as COCO for object detection and ImageNet for ResNet) made the training process more efficient and allowed the model to perform better by leveraging knowledge from large, diverse datasets.

What are the team's plans for future improvements on this project?

Future improvements include acquiring more up-to-date and comprehensive datasets, improving model generalization for real-world conditions, exploring real-time detection capabilities, and experimenting with video data and more efficient training methods.