acceptodds
Under review as a conference paper at ICLR 2027

GeoBias: Geometric Bias Correction via Decision Boundary Translation for Long-Tailed Recognition

Abstract

Real-world visual recognition tasks often exhibit long-tailed label distributions, where a few head classes dominate training while most tail classes are underrepresented. Standard empirical risk minimization biases classifiers toward head classes, severely degrading tail-class generalization. Existing decoupling methods mainly adjust classifier weights, overlooking the bias term as an independent lever for decision-boundary correction. We propose GeoBias, a geometry-driven bias correction method that fixes the learned weight directions and analytically computes bias shifts from class centroids and pairwise boundary geometry. Multi-class conflicts are resolved by weighted aggregation, and a time-decaying anchoring regularizer prevents SGD from eroding the corrected biases. GeoBias is single-stage and plug-and-play, requiring no data re-sampling, architectural modification, or complex training schedule. Experiments on CIFAR-10-LT, CIFAR-100-LT, and ImageNet-LT show that GeoBias consistently improves tail-class accuracy and achieves competitive or superior performance relative to single-stage and two-stage methods. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long-tailed recognition, Class imbalance, Decision boundary correction, Bias term adjustment, Geometric calibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_cgPsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tackles long-tailed recognition by changing only the bias term of the classifier, instead of retraining the weights as decoupling methods do. During normal CE training the method periodically computes class centers. For each pair of classes it shifts the bias so that the decision boundary moves toward the head class, based on the head-class samples that sit close to the boundary. These pairwise shifts are averaged into one adjustment per class, and a regularizer keeps the bias from drifting between calibrations. Experiments on CIFAR-10-LT, CIFAR-100-LT and ImageNet-LT against eight baselines are used to support three claims, that decoupling methods overlook the bias term, that GeoBias beats single-stage methods and matches two-stage ones, and that it consistently improves tail accuracy.

Critical strengths
  • Simplicity and Clarity: The core geometric intuition is straightforward. The paper clearly and correctly illustrates how shifting the bias translates the decision boundary along the center line between two class centroids without altering the learned weight directions.

  • Ease of Integration: The method is highly plug-and-play. It integrates easily into existing training pipelines, utilizing a standard linear classifier without requiring weight normalization, data resampling, or a complex two-stage training paradigm.

Critical weaknesses
  • Flawed novelty claims regarding bias: The paper says decoupling methods overlook the bias term as a separate lever for moving the boundary, but DisAlign, which the paper cites, learns an extra bias for every class, and cRT retrains both the weights and the bias. Outside decoupling, post-hoc logit adjustment and Balanced Softmax can also be seen as per-class bias corrections, which is relevant because the conclusion extends the claim to existing methods in general. As I read it, the new part is how the bias is estimated, from the geometry of each pair of classes, and that this estimate is applied periodically during training.

  • Absence of crucial baselines: The paper does not compare with the logit adjustment of Menon et al., in its post-hoc or loss form, or with Balanced Softmax, which it does not cite. There is also no ablation that replaces the geometric target with the log ratio of the two class sizes. Published numbers suggest this comparison may not be easy, at least against Balanced Softmax. The SADE paper uses a cosine classifier for all methods. Its Softmax numbers on CIFAR-100-LT are within half a point of this paper's CE, and its Balanced Softmax numbers are higher than this paper's GeoBias numbers in all three settings it reports. The setups differ, so this is only a rough indication, but I think this is the comparison that would most strengthen the paper.

  • Substandard reproduction of baselines: Some reproduced baselines are well below their published results, which makes the comparison harder to interpret. GCL gets 39.90 on ImageNet-LT here against 54.88 in its own paper, and it is listed as single-stage although its paper trains in two stages. cRT ends up below CE on ImageNet-LT, while Kang et al. report cRT 5.7 points above joint training with the same ResNet-50 and 90-epoch setup. Separately, the claim in Table 4 that a tuned GeoBias beats the best two-stage method is based on configurations picked on the test set, and the margins over LWS are 0.23 and 0.50.

  • Severe contradictions between text and tables: A few statements in the text do not quite match the tables. On CIFAR-100-LT with ratio 200, the paper says GeoBias more than doubles GCL's tail accuracy, but 8.85 against 5.75 is about 1.5 times. On CIFAR-10-LT it is said to beat all single-stage methods consistently, while at ratio 10 it is slightly below CE, ALA and GCL. On ImageNet-LT it is said to beat GCL in overall accuracy, but GCL scores 39.90 against 39.27. The ablation reading also seems off. In Table 4, removing the anchoring regularizer improves both top-1 and tail accuracy at both ratios, whereas the text describes the regularizer as crucial for tail accuracy. I would suggest revising these sentences to match the numbers.

Questions for authors
  • What happens if the geometric target is replaced by the log ratio of the two class sizes, or if the bias is calibrated only once after training?

  • Was GCL run with only its first stage? What second-stage settings did you use for cRT and LWS, and do you have an idea why cRT is below CE on ImageNet-LT?

  • How were the default hyperparameters chosen? The paper describes no validation split apart from the test data, so I would suggest choosing both the defaults and the Table 4 settings on a validation split taken from the training set.

  • The paper mentions high training efficiency, but I could not find training times. How long does one calibration take on ImageNet-LT, where it loops over every pair of classes?

Reviewer_d1hXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes GeoBias a single stage logit-adjustment method for long-tailed classification. The logit adjustment is inspired by feature geometry and the classifier decision boundaries. It periodically updates the classifier margins during training based on intra-class compactness and weighted global least squares distance. It also uses a bias anchoring mechanism that regularizes the margin adjustments. The paper tests GeoBias using three datasets, CIFAR10 CIFAR 100 and ImageNet-LT using imbalance ratios 10,50,100,200 and the Imagenet-LT native imbalance ratio. It uses ResNet 32 for Cifar and Resnet 50 for Imagenet-LT and shows increased performance compared to standard logit adjustment methods such as ALA,LDAM, LWS, GCL and two stage methods like cRT. It conducts ablations on the frequency of margin-readaptation during training and the effects of bias regularization.

Critical strengths

The paper has a solid theoretical framework. The GeoBias is a novel instantiation of logit-adjustment method family. It is a nice property to be able to analytically and dynamically control the bias adjustment based on feature compactness and the classifier margins.

Critical weaknesses

I think that the paper could use some more figures to visually illustrate some ideas about some abstract mathematical concepts, like LDAM did. Another idea would be to give more visualization in the experiments to show how the top-1 accuracy changes as we change the margins for head an tail classes. My most critical concern is about the performance. I think that the current baseline is too low, in the same status as 6-7 years ago. There have been many works that considerably pushed the performance on ImageNet-LT, CIFAR Places-LT and iNaturalist-18 and this paper does not cite or compare against. The other issue is that the estimated feature compactness/ classifier margins and so on, are all estimated from a model that is trained on long-tail dataset and for rare classes these estimates are noisy. Thus I think using a better backbone like Resnext50 or using some data augmentation or more advanced techniques then the estimated qualities are less noisy and it gives GeoBias a better platform to be tested. Therefore I suggest applying GeoBias to more advanced techniques, there are many bag-of-trick papers from long tail and even methods from balanced ImageNet feature learning can be adapted easily for long-tail learning.

Questions for authors

Did you try the GeoBias under two stage learning? Could we apply GeoBias to iNaturalist18 and Places-LT? Does it work with ViT-B under MAE pretraining? How does it compare against SAM-based methods?

In general, I like the paper and I would raise my score if the authors do more experiments and further push the performance using more modern techniques. To understand my performance request specifically please see these two papers [a1], [a2], [a3] they sould be taken as a guide I understand that direct comparison between them might not be appropriate but they set a high bar for performance and they authors can adapt some of these techniques:

[a1]Xu et. al. Learning Imbalanced Data with Vision Transformers, in CVPR2023 [a2]Rangwani et. al. DeiT-LT Distillation Strikes Back for Vision Transformer Training on Long-Tailed Datasets, in CVPR2024 [a3] Alexandridis et al. Adaptive Parametric activation in ECCV2024

Reviewer_135NOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GeoBias, a geometry-driven classifier bias correction method for long-tailed visual recognition. While existing decoupling methods primarily adjust weight matrices or rely on static logit adjustments, they overlook the classifier bias term as an independent lever for correcting the decision boundary, which the proposed approach focuses on. The paper proposes a single-stage, plug-and-play framework that fixes learned weight directions and analytically computes bias shifts based on class centroids and pairwise boundary geometry. Multi-class conflicts are resolved using a weighted global aggregation scheme, and a time-decaying bias-anchoring regularizer prevents SGD from eroding the corrections.

Critical strengths

Explicitly isolating the bias term as an independent tool for decision-boundary translation avoids disrupting the already-learned discriminative feature-weight directions.

GeoBias integrates directly into standard single-stage training with minimal overhead

Critical weaknesses

z(x) is not defined anywhere.

The meaning of Head/Medium/Tail is not described anywhere.

The methods being compared are at least 4 years old. Are there no works in this field from the last 4 years?

The proposed method seems more like a basic training hack.

Questions for authors

Definition of z(x)

The evaluation tables report performance breakdowns across Head, Medium, and Tail classes, yet the paper omits the explicit sample-size boundaries or proportion thresholds defining these splits.

How does GeoBias compare against more recent, contemporary long-tailed recognition methods?

Since GeoBias performs periodic full-forward passes over the entire training set during geometric calibrations to compute class centroids and SPD scores, what is the exact wall-clock training overhead compared to standard single-stage empirical risk minimization or baseline decoupling methods?

Reviewer_7sMMOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoBias for long-tailed recognition. It corrects classifier biases using class centroids and pairwise boundary geometry. The corrections are computed from correctly classified head-class samples near each boundary. They are then combined through weighted averaging. A decaying anchoring regularizer constrains subsequent bias updates. Experiments are conducted on CIFAR-10-LT, CIFAR-100-LT, and ImageNet-LT.

Critical strengths

The calibration step is simple and interpretable, which shifts pairwise boundaries by changing classifier biases and does not requires additional network modules or re-sampling.

On CIFAR-10-LT at the imbalance ratio of 200, GeoBias improves overall accuracy over GCL from 72.90% to 76.03% and tail accuracy from 51.75% to 71.90%.

Critical weaknesses
  1. The benefit of the geometric correction is somewhat confusing. Eq.19 subtracts the mean logit margin of selected head-class samples. How does GeoBias compare with Logit Adjustment and Balanced Softmax under matched settings? An ablation of sample selection and pairwise weighting would help identify where the gains come from.

  2. The choice of correction size needs more justification. The update sets the average margin of the selected head samples to zero. Would a smaller shift preserve more head accuracy while retaining the tail gains? It is also unclear how the aggregation handles pairs with qijq_ij close to zero.

  3. Table 4 does not support the claimed benefit of anchoring. Removing it improves both overall and tail accuracy at both imbalance ratios. At IR 200, tail accuracy rises from 8.85% to 11.90%. Does anchoring reduce variation across random seeds or stabilize the training trajectory?

  4. Some conclusions overstate the results. On ImageNet-LT, GeoBias trails GCL in overall accuracy (39.27% vs. 39.90%). The reported resules are much lower than the original resluts in GCL.

  5. The compared baselines are out of date. There is no recent SOTA works.

Questions for authors

Please refer to the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.