- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces Open-Domain FSCIL, in which the base classes come from a single real-image domain but the few-shot samples of new classes may come from mixed, unknown domains. The proposed LeVa builds on CLIP. It transfers intra-class variations learned from base classes to new classes, keeps the variations that match a predicted class distribution, and then uses them to model variation around the text anchors for recognition. The authors also build a new benchmark, Hybrid-miniImageNet, and report gains on it, on DomainNet and on standard FSCIL benchmarks.
Critical strengths
- Relaxing the single-domain assumption of FSCIL is practically relevant, and the new benchmark is a useful contribution. Moving these variations between classes by parallel transport on the hypersphere is an elegant and novel design.
- LeVa outperforms the strongest baseline on both open-domain benchmarks and also does well in the standard setting. The ablations cover the main components.
Critical weaknesses
- Most baselines are ResNets trained from scratch, compared against pretrained CLIP. Only one CLIP-based method (BiMC) is included. Zero-shot CLIP (85.5 AHM) and relevant pretrained-model FSCIL methods (e.g. Goswami et al., 2024; FSHIL, CVPR 2026) are missing from the main tables, so it is hard to tell how much of the gain comes from the method itself.
- The distribution-guided selection, the main novelty, beats random variation transfer by only 0.9 AHM.
- Key parts of Variation Recovery (e.g. the descriptor q(S, t), the kernel and RFF settings, the memory size) are not specified. The construction of Hybrid-miniImageNet is also unclear, including the image sources per domain and the test-set composition.
Questions for authors
- How exactly is Hybrid-miniImageNet built (image sources and counts per domain, test-set composition), and will it be released?
- Could you add zero-shot CLIP, a simple trained CLIP baseline and one or two pretrained-model FSCIL methods to Tables 1–2?
- How much does the learned distribution predictor help compared with simply using the support mean?
6 comments
Response to Q2 and W1: Comparisons with pretrained-model baselines
We thank the reviewer for highlighting the need for stronger pretrained-model comparisons. We have expanded the evaluation on both benchmarks to include zero-shot CLIP, CoOp as a trainable CLIP baseline, and two pretrained-model FSCIL methods: FSCIL-Calibration, suggested by the reviewer, and HyCal. We will incorporate these results into Tables 1–2 of the revised manuscript.
For the suggested FSHIL method, we were unable to obtain its official code and data splits at the time of our experiments and therefore do not report results for it. We instead include HyCal as an additional pretrained-model FSCIL comparison. For HyCal, we used grid search to select on Hybrid-miniImageNet and on DomainNet.
As shown below, LeVa achieves the highest HM in every incremental session on both benchmarks. Among the added baselines, CoOp achieves the strongest AHM, demonstrating the importance of comparing against a trainable CLIP model. LeVa nevertheless exceeds CoOp by 1.84 percentage points on Hybrid-miniImageNet and 4.42 percentage points on DomainNet in AHM. Together with the controlled component ablations, these comparisons provide evidence of gains beyond simply adopting a pretrained CLIP backbone.
Table R1. Additional baseline comparisons on Hybrid-miniImageNet. S0 reports base-session accuracy (%); S1–S8 report HM (%). AHM is averaged over S1–S8.
| Method | S0 Acc. | S1 HM | S2 HM | S3 HM | S4 HM | S5 HM | S6 HM | S7 HM | S8 HM | AHM |
|---|---|---|---|---|---|---|---|---|---|---|
| Zero-shot CLIP | 91.80 | 91.10 | 87.86 | 87.45 | 84.40 | 84.36 | 82.61 | 83.03 | 83.26 | 85.51 |
| CoOp | 96.40 | 94.32 | 91.67 | 91.26 | 88.62 | 88.77 | 86.85 | 86.62 | 87.07 | 89.40 |
| FSCIL-Calibration | 97.87 | 77.83 | 70.75 | 72.99 | 67.83 | 58.97 | 58.28 | 59.35 | 60.44 | 65.81 |
| HyCal | 94.38 | 89.61 | 87.83 | 85.95 | 81.86 | 81.60 | 77.79 | 76.97 | 76.75 | 82.29 |
| LeVa | 96.95 | 95.76 | 93.92 | 93.13 | 90.50 | 90.49 | 88.56 | 88.52 | 89.04 | 91.24 |
Table R2. Additional baseline comparisons on DomainNet. S0 reports base-session accuracy (%); S1–S8 report HM (%). AHM is averaged over S1–S8.
| Method | S0 Acc. | S1 HM | S2 HM | S3 HM | S4 HM | S5 HM | S6 HM | S7 HM | S8 HM | AHM |
|---|---|---|---|---|---|---|---|---|---|---|
| Zero-shot CLIP | 67.08 | 67.00 | 66.56 | 65.15 | 63.90 | 62.49 | 62.55 | 62.31 | 62.43 | 64.05 |
| CoOp | 78.77 | 73.32 | 71.16 | 69.45 | 67.40 | 66.53 | 67.04 | 66.24 | 66.56 | 68.46 |
| FSCIL-Calibration | 81.80 | 51.55 | 50.31 | 46.24 | 44.38 | 44.16 | 44.15 | 41.99 | 44.22 | 45.87 |
| HyCal | 82.79 | 63.05 | 58.56 | 56.86 | 55.77 | 53.59 | 52.46 | 49.65 | 48.78 | 54.84 |
| LeVa | 87.01 | 78.46 | 76.08 | 73.74 | 72.25 | 70.98 | 71.13 | 70.15 | 70.22 | 72.88 |
- Zero-shot CLIP — Learning Transferable Visual Models From Natural Language Supervision [ICML 2021].
- CoOp — Learning to Prompt for Vision-Language Models [IJCV 2022].
- FSCIL-Calibration — Calibrating Higher-Order Statistics for Few-Shot Class-Incremental Learning with Pre-trained Vision Transformers [CVPR Workshops 2024].
- HyCal — HyCal: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning [CVPR 2026]
Response to W2 and Q3: Comparisons with random selection and the support-based reference
We thank the reviewer for suggesting a closer examination of distribution-guided recovery. We acknowledge that the AHM gain over random variation selection on Hybrid-miniImageNet is modest (0.91 percentage points). We extended this comparison to DomainNet and additionally evaluated a support-based distributional reference.
(1) Selection using the learned distributional reference versus random selection. The random baseline samples from the same semantically filtered candidate pool, retains the same number of recovered features, and uses identical downstream processing. Selection guided by the learned distributional reference improves AHM by 0.91 percentage points on Hybrid-miniImageNet and 4.11 percentage points on DomainNet, with a 5.86-percentage-point gain in average novel-class accuracy on DomainNet. On Hybrid-miniImageNet, random selection already achieves 90.33% AHM, leaving relatively limited headroom for further gains. On DomainNet, which includes additional domains such as Infograph and Quickdraw, the random baseline achieves 68.77% AHM. These more challenging domain conditions may provide greater scope for the learned distributional reference to improve candidate selection.
(2) Learned versus support-based reference. We replace the predicted distributional reference with the empirical kernel mean of the support samples, , while keeping the remaining components unchanged. The learned reference improves average novel-class accuracy by 4.80 percentage points and AHM by 2.39 percentage points, with a 1.34-percentage-point decrease in average base-class accuracy. This comparison demonstrates the additional benefit of predicting a distributional reference beyond directly using the observed supports.
Table R3. Variation recovery ablations on DomainNet.
| Variant | Avg. Novel ↑ | Avg. Base ↑ | AHM ↑ |
|---|---|---|---|
| Random selection | 60.21 | 80.98 | 68.77 |
| Support-based reference | 61.27 | 83.42 | 70.49 |
| Learned reference (LeVa) | 66.07 | 82.08 | 72.88 |
These controlled comparisons support the contributions of both distribution-guided selection and the learned distributional reference.
Thank you for your thoughtful and encouraging feedback.
Response to Q1 and W3: Dataset construction and implementation details.
We thank the reviewer for pointing out these omissions and apologize for the lack of clarity in the original manuscript. We provide the missing dataset construction and implementation details below and will incorporate them into the revised manuscript.
(1) Construction and release of Hybrid-miniImageNet. Hybrid-miniImageNet preserves the original miniImageNet class split and base-class training and test data. For novel classes, Real-domain images are sampled from the original miniImageNet novel-class set, while stylized images are primarily sourced from ImageNet-R (Rendition) and ImageNet-Sketch, supplemented with web-collected images for classes insufficiently covered by these datasets. Each novel class contains 600 images across four domains—Real, Clipart, Painting, and Sketch—with 150 images per domain. We reserve 25 images per domain for testing, yielding 100 test images per novel class and 4,000 test images across all 40 novel classes. The remaining 125 images per domain constitute the training pool from which the five labeled supports per novel class are sampled. We will publicly release the dataset.
(2) Descriptor q(Se,tce).
We define the descriptor as
q(Se,tce)=[μ(Se);r(Se);tce],
where semicolons denote concatenation. Here, μ(Se) summarizes the support distribution through its empirical kernel mean, r(Se) characterizes the pairwise visual similarity structure among support samples, and tce provides the class text representation.
(3) Kernel and random Fourier features.
We use a Gaussian RBF kernel on 64-dimensional PCA projections of CLIP features. The kernel bandwidth σ is set to the median pairwise Euclidean distance between base-training features after PCA projection. We approximate the kernel using 256-dimensional random Fourier features (RFF), with frequencies sampled from N(0,σ−2I) and phases sampled uniformly from [0,2π). The PCA projection, kernel bandwidth, and sampled RFF parameters are determined during the base stage and remain fixed throughout subsequent incremental sessions.
(4) Variation memory.
We maintain a feature-level variation memory. Each class stores one prototype, while each variation entry stores a tangent vector linked to its corresponding class center through a class index.
On Hybrid-miniImageNet, the base memory contains 9,600 variation entries, with 160 entries per base class. Each novel class adds five entries from its observed supports. After the final session, the memory contains 9,800 variation entries and 100 prototypes, requiring approximately 20.4 MB.