- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes Spatial Moment for long-tailed recognition. The main idea is to use different pooling orders for different experts. Head-focused experts use more global information, while tail-focused experts focus more on local features. The method also uses a logit correction term and a regularization loss to improve classification. At test time, the experts are combined based on how confident their predictions are. Experiments on several long-tailed datasets show that the method achieves competitive results and improves tail-class accuracy.
Critical strengths
- The paper presents an interesting idea of using different pooling orders for different experts in long-tailed recognition.
- The experiments cover several long-tailed datasets and show competitive results, especially for tail classes.
Critical weaknesses
-
The core motivation is not well supported. The paper assumes that head classes rely more on global context, while tail classes rely more on local features. However, class frequency does not necessarily determine the spatial scale of useful features. The paper does not provide direct evidence to support this assumption.
-
The link between high activations and useful features is unclear. GeM pooling focuses on regions with high activations. However, these regions may contain background noise or spurious features rather than useful information. The related work Shortcut-Resistant CAM Distillation for Long-Tailed Recognition (ICML 2026) has studied similar issues in long-tailed recognition. The paper does not discuss this work in enough detail or clearly explain how its method differs from existing methods in handling background bias and spurious correlations.
-
The need for preference-conditioned modulation is unclear. Table 5 shows that a fixed GeM ladder (1, 2, 3) already achieves 53.4% overall accuracy, while the proposed modulation only increases it to 53.8%. This suggests that most of the gains come from using different pooling orders across experts, rather than from preference modulation. It is unclear whether the added complexity is necessary.
-
Limited novelty and too many hand-tuned settings. The method mainly combines existing techniques, such as GeM pooling, cosine classifiers, logit calibration, margin losses, and confidence-based expert weighting. It also uses many manually chosen hyperparameters, including the modulation coefficient, Dirichlet concentration, radial correction strength, margin coefficient, regularization weight, and expert weighting ratio. The paper provides limited analysis of these choices. Overall, the method seems more like a combination of existing techniques than a new approach that offers deeper insights into long-tailed recognition.
Questions for authors
See weaknesses.
2 comments
Thank you for your response. However, my main concerns remain. The new experiments only change pooling orders in a frozen model and do not fully support the core motivation. Also, high activations do not necessarily indicate useful features, and the need for preference modulation remains unclear given the strong results of the fixed GeM ladder. The concerns about limited novelty and many manually chosen hyperparameters are also not fully addressed. Therefore, I will keep my original score.
Response to Reviewer HK1g
Thank you for the comments on motivation, shortcuts, and complexity. Our supplementary analyses focus on the architectural question: whether spatial aggregation can usefully participate in expert specialization before classification.
1. Frequency objectives and aggregation sensitivity
The design asks whether experts with different frequency objectives benefit from different aggregation sensitivities. This formulation accommodates varied object sizes and spatial layouts; frequency defines the learning objective rather than a fixed assumption about object geometry.
In a separate supplementary CIFAR-100-LT IF=100 evaluation—ResNet-32, 200 epochs, seed 1—we froze the model and compared its deployed orders with uniform GAP, uniform p=2, uniform p=3, fixed (1,2,3), and reversed deployed orders.
The deployed configuration achieves 53.04% Overall / 32.80% Few. The five alternatives range from 50.06% to 52.77% Overall and 31.47% to 32.47% Few. Reversing the deployed orders gives 52.34% Overall / 32.03% Few.
These controls directly establish that aggregation and expert–order assignment affect recognition in the trained model. They support studying spatial aggregation as an architectural variable, with the comparison scoped to fixed-checkpoint interventions.
2. High activations and background shortcuts
Activation magnitude and foreground relevance are distinct properties. GeM provides explicit control over how strongly different activations contribute to the pooled representation, and the new controls show that this choice affects recognition.
SRCD addresses evidence selection through teacher maps constructed from head-proxy CAMs and explicit distillation. Spatial Moment operates through aggregation before classification, using no such CAM supervision.
This distinction clarifies the scope of our evidence: the measurements support control of activation concentration and its effect on recognition. Semantic localization and shortcut resistance remain separate properties from concentration itself.
3. Conditioning and deployment complexity
A fixed ladder implements the spatial hierarchy, while shared conditioning supplies bounded adjustments and connects pooling to classifier generation during training. At deployment, q⋆=softmax(a) is fixed, so classifier weights and pooling orders can be precomputed and cached; consensus weighting remains sample-dependent. This fixed-parameter deployment form follows directly from the formulation.
4. Novelty and hyperparameters
GeM, cosine classifiers, and class-aware losses provide established building blocks. Our contribution is to introduce expert diversity at the spatial aggregation stage and study its effect through an explicit sensitivity mechanism and controlled comparisons. The pooling branch reuses the existing preference coordinates and adds no independent learnable pooling parameters.
The supplementary pooling comparisons hold all auxiliary coefficients fixed, directly testing aggregation changes without joint retuning. A matched-training-recipe comparison also gives 53.04% versus 52.04% Overall, and 32.80% versus 31.17% Few, for Spatial Moment and its reference baseline.
We position the auxiliary components as supporting design choices. The central contribution is the concrete architectural insight of spatial expert specialization, supported by the observed aggregation behavior and recognition benefits.
Sincerely,
The Authors