Inverse Image Frequency for Long-Tailed Image Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world data often exhibits long-tailed distributions, leading to imbalanced datasets.
• Machine learning models trained on imbalanced data perform well on frequent categories but poorly on rare ones, causing biased predictions.
• This imbalance degrades overall model performance and limits applicability in practical scenarios.

Purpose of the Study:

• To propose a novel de-biasing method, Inverse Image Frequency (IIF), to address the challenges posed by long-tailed distributions in image datasets.
• To improve the performance of machine learning models, particularly in recognizing under-represented categories.
• To reduce false positive detections in downstream tasks like instance segmentation.

Main Methods:

• Introduced Inverse Image Frequency (IIF), a de-biasing technique.
• IIF applies a multiplicative margin adjustment to the logits in the classification layer of convolutional neural networks.
• The method was evaluated on various long-tailed benchmarks including ImageNet-LT, CIFAR-LT, Places-LT, and LVIS.

Main Results:

• IIF demonstrated superior performance compared to existing methods on multiple long-tailed benchmarks.
• Achieved 55.8% top-1 accuracy on ImageNet-LT using ResNet50.
• Reached 26.3% segmentation AP on LVIS using MaskRCNN ResNet50, with fewer false positive detections.

Conclusions:

• Inverse Image Frequency (IIF) effectively mitigates bias in machine learning models trained on imbalanced, long-tailed data.
• The proposed method shows significant improvements over state-of-the-art techniques.
• IIF is particularly beneficial for downstream tasks such as long-tailed instance segmentation, enhancing detection accuracy.