Fractal Calibration for long-tailed object detection
Konstantinos Panagiotis Alexandridis, Ismail Elezi, Jiankang Deng, Anh, Nguyen, Shan Luo
TL;DR
FRACAL is a post-calibration method that uses fractal dimension to improve long-tailed object detection by balancing class probabilities without retraining, enhancing rare class detection across multiple datasets.
Contribution
It introduces a novel fractal dimension-based logit adjustment for post-calibration in long-tailed detection, outperforming previous methods without additional training.
Findings
Boosts rare class detection by up to 8.6%.
Surpasses previous methods on LVIS dataset.
Generalizes well to COCO, V3Det, and OpenImages.
Abstract
Real-world datasets follow an imbalanced distribution, which poses significant challenges in rare-category object detection. Recent studies tackle this problem by developing re-weighting and re-sampling methods, that utilise the class frequencies of the dataset. However, these techniques focus solely on the frequency statistics and ignore the distribution of the classes in image space, missing important information. In contrast to them, we propose FRActal CALibration (FRACAL): a novel post-calibration method for long-tailed object detection. FRACAL devises a logit adjustment method that utilises the fractal dimension to estimate how uniformly classes are distributed in image space. During inference, it uses the fractal dimension to inversely downweight the probabilities of uniformly spaced class predictions achieving balance in two axes: between frequent and rare categories, and between…
Peer Reviews
Decision·ICLR 2025 Conference Withdrawn Submission
The paper tackles a relevant and important problem. Post-training calibration of object detectors is of practical importance. Integrating spatial location awareness into logit adjustment is a novel and interesting idea. The authors report improved results on many datasets.
There are three major weaknesses: 1. Clarity: The paper is not well-written. Important pieces of information are not clear. 2. Substantiation of the main claim: I am not sure whether the main claim of the paper, which is incorporating spatial location distributions, is substantiated. 3. Performance comparisons: Experimental comparisons do not provide a clear indication whether we should use this technique with modern object detectors. I explain them in detail below. ## Clarity: - Equat
1. This paper addresses an underexplored aspect in long-tailed object detection by introducing spatial information into post-calibration using fractal dimensions. This is innovative and may inspire new directions in handling data imbalance in object detection. 2. The proposed FRACAL method achieves significant improvements in performance for rare classes without retraining or additional inference costs, making it a practical tool for applications needing efficient solutions. 3. The authors con
1. While the fractal dimension is a core component of the proposed method, the paper could benefit from a more detailed analysis on how different fractal dimension values influence performance across categories. This would add insight into the method's adaptability to datasets with varying spatial distributions. 2. The paper briefly discusses the limitations of grid-based approaches but does not thoroughly compare FRACAL against alternative spatial calibration methods. Including this would help
1) This paper attempts to introduce the class-location dependency to post-calibration, which enhances the performance of rare categories. 2) This work fuses the space information by integrating the fractal dimension and optimizes the detection performance. 3) This is a plug-in-plug-out module, so it does not need extra training.
1) I feel confused about the description of Fig. 1. Why class prior ps(y) is used to align ps(y, u|x) with pt(y|x) rather than align ps(y|x) with pt(y|x)? Do you mean that space information has been considered by previous work just for training data? 2) Although the fractal dimension is introduced, the theoretical explanation of how the fractal dimension specifically affects the detection of rare categories can be more detailed. It is possible to further discuss how the fractal dimension improv
+ Calibration is a viable method for reducing the impact of class/data imbalance. + The gap identified by the authors (i.e., taking location distribution into account during calibration) is definitely promising. + The proposed calibration method taking into account prior distribution of samples across space as well as the fractal dimension is novel.
1. Although the proposed approach is novel, some of the individual steps are not sufficiently motivated. For example: 1.1. "Eq. 9 downweighs the classes that appear most uniformly and it upweighs the classes that appear less uniformly. In practice, this scheme enforces a centre bias for frequent classes and no bias for rare classes, as shown in Fig. 2-bottom-right. Intuitively, removing the bias from the rare classes is better than rectifying it because it produces balanced detectors and aligns
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsCurrency Recognition and Detection · Industrial Vision Systems and Defect Detection · Image Retrieval and Classification Techniques
MethodsFocus