[Paper Review] Resolving Class Imbalance in Object Detection with Weighted Cross Entropy Losses

Trong Huy Phan, Kazuma Yamamoto|arXiv (Cornell University)|Jun 2, 2020

Advanced Neural Network Applications19 references54 citations

TL;DR

The paper investigates how weighted cross-entropy losses can mitigate class imbalance in object detection, comparing Balanced Cross Entropy, Focal Loss, and Class-Balanced Loss on imbalanced datasets like BDD100K.

ABSTRACT

Object detection is an important task in computer vision which serves a lot of real-world applications such as autonomous driving, surveillance and robotics. Along with the rapid thrive of large-scale data, numerous state-of-the-art generalized object detectors (e.g. Faster R-CNN, YOLO, SSD) were developed in the past decade. Despite continual efforts in model modification and improvement in training strategies to boost detection accuracy, there are still limitations in performance of detectors when it comes to specialized datasets with uneven object class distributions. This originates from the common usage of Cross Entropy loss function for object classification sub-task that simply ignores the frequency of appearance of object class during training, and thus results in lower accuracies for object classes with fewer number of samples. Class-imbalance in general machine learning has been widely studied, however, little attention has been paid on the subject of object detection. In this paper, we propose to explore and overcome such problem by application of several weighted variants of Cross Entropy loss, for examples Balanced Cross Entropy, Focal Loss and Class-Balanced Loss Based on Effective Number of Samples to our object detector. Experiments with BDD100K (a highly class-imbalanced driving database acquired from on-vehicle cameras capturing mostly Car-class objects and other minority object classes such as Bus, Person and Motor) have proven better class-wise performances of detector trained with the afore-mentioned loss functions.

Motivation & Objective

• Motivate the problem of class imbalance in object detection and its impact on detector performance.
• Evaluate how weighted cross-entropy variants can improve minority class accuracy in detectors.
• Demonstrate effectiveness on a highly imbalanced driving dataset (BDD100K).

Proposed method

• Apply Weighted Cross Entropy variants (Balanced Cross Entropy, Focal Loss, Class-Balanced Loss Based on Effective Number of Samples) to object detectors.
• Use these losses to train detectors on imbalanced data to assess class-wise performance improvements.
• Experiment with a driving dataset (BDD100K) exhibiting majority Car objects and minority classes like Bus, Person, Motor.
• Compare detector performance across loss variants to identify gains for minority classes.

Experimental results

Research questions

• RQ1 Can weighted cross-entropy losses improve per-class detection accuracy on imbalanced datasets?
• RQ2 Which loss variant most effectively boosts minority class performance in object detection?
• RQ3 Do these losses affect overall detector performance or only minority classes on driving datasets?
• RQ4 How do results on BDD100K validate the approach for real-world imbalanced scenarios?

Key findings

• Weighted Cross Entropy variants show improved class-wise performance for minority object classes.
• Experiments on BDD100K demonstrate gains for less frequent classes such as Bus, Person, and Motor without stating exact figures.
• The paper provides evidence that balancing/weighting strategies can help detectors trained with standard architectures.
• Among the losses, the study identifies Focal Loss and Class-Balanced Loss as effective options in imbalanced settings.