A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object Detection (2009.13592v4)

Abstract: We propose average Localisation-Recall-Precision (aLRP), a unified, bounded, balanced and ranking-based loss function for both classification and localisation tasks in object detection. aLRP extends the Localisation-Recall-Precision (LRP) performance metric (Oksuz et al., 2018) inspired from how Average Precision (AP) Loss extends precision to a ranking-based loss function for classification (Chen et al., 2020). aLRP has the following distinct advantages: (i) aLRP is the first ranking-based loss function for both classification and localisation tasks. (ii) Thanks to using ranking for both tasks, aLRP naturally enforces high-quality localisation for high-precision classification. (iii) aLRP provides provable balance between positives and negatives. (iv) Compared to on average $\sim$6 hyperparameters in the loss functions of state-of-the-art detectors, aLRP Loss has only one hyperparameter, which we did not tune in practice. On the COCO dataset, aLRP Loss improves its ranking-based predecessor, AP Loss, up to around $5$ AP points, achieves $48.9$ AP without test time augmentation and outperforms all one-stage detectors. Code available at: https://github.com/kemaloksuz/aLRPLoss .

• The paper introduces the aLRP Loss function that unifies classification and localisation through a ranking-based framework.
• It streamlines model tuning by reducing hyperparameter complexity to a single parameter for step-function approximation.
• It achieves up to a 5.4 AP improvement on the COCO dataset with a 48.9 AP score, outperforming traditional one-stage detectors.

A Ranking-Based, Balanced Loss Function Unifying Classification and Localisation in Object Detection

The paper introduces a novel approach to object detection, presenting the average Localisation-Recall-Precision (aLRP) Loss function. This loss function is noteworthy for its unification of classification and localisation tasks under a single, ranking-based framework. Object detection traditionally involves balancing classification and localisation objectives, commonly combined with a hyperparameter weighing their contributions. However, this method faces challenges, such as the lack of task correlation, hyperparameter tuning, and imbalances in training data. The authors propose aLRP Loss, which addresses these challenges by extending the Localisation-Recall-Precision metric into a loss function suitable for both classification and localisation.

Key Contributions and Methodology

The aLRP Loss function is defined as the average of LRP values over positives on the Recall-Precision curve, drawing inspiration from how AP Loss was derived. The following are the standout contributions of the paper:

1. Unified Ranking-Based Framework: aLRP Loss incorporates ranking into both classification and localisation processes. This results in a natural enforcement of high-quality localisation for high-precision classifications, a correlation the authors argue was missing in previous methods.
2. Minimal Hyperparameters: Unlike traditional loss functions, which may employ around 6 hyperparameters requires adjusting through tedious experiments, aLRP Loss needs only one hyperparameter related to its step-function approximation. In practice, the authors reported not requiring tuning this hyperparameter.
3. Balanced Training: The paper presents a theoretical foundation showing that ranking-based loss functions ensure balance between positive and negative samples during training using an error-driven optimization strategy. This aspect is showcased through mathematical formulations and proofs aligning with perceptron learning principles.
4. Performance Enhancement: On the COCO dataset, the aLRP Loss achieves an improvement of up to 5.4 AP over AP Loss. Without test-time augmentations, it achieves a 48.9 AP score, thereby outperforming contemporary one-stage detectors.

To compute aLRP, errors are inferred based on rankings of both classification scores and localisation outputs, thereby naturally linking the two tasks. The aLRP Loss is optimized using a generalized error-driven update method that distributes losses over the false positive predictions and correlates them with localisation errors through meaningful ranking mechanisms.

Implications and Future Directions

The implications of applying aLRP Loss are twofold. Practically, the loss function's ability to match state-of-the-art detection performances while simplifying hyperparameter tuning can lead to easier deployment of object detection models across various applications. Theoretically, this work challenges the conventional separation between classification and localisation tasks by proposing a correlated approach, which could inspire further research on unified objectives in multi-task learning scenarios.

One area for further exploration is expanding the application of aLRP Loss beyond object detection. Given its foundation in ranking-based metrics, similar methodologies could be developed for tasks like instance segmentation or panoptic segmentation, where multiple objectives must be optimized concurrently.

In conclusion, this paper provides a significant advancement in loss function design for object detection by bridging classification and localisation tasks, reducing the need for manual tuning, and achieving high performance, thus contributing to both the theoretical understanding and practical application of object detection technologies.