- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
STRA adapts frozen CLIP models online without backpropagation or raw-feature caching. It estimates the stream’s uncentered second moment, applies a regularized inverse-square-root map to image features and text-anchored class prototypes, and normalizes both before scoring. Sparse pseudo-labels from frozen CLIP update the prototypes. A pullback-vMF model interprets the classifier. Evaluation covers natural shifts, ten recognition datasets, and fifteen corruption types using RN50 and ViT-B/16.
Critical strengths
-
Geometric alignment and consistent gains: Normalizing transformed features and prototypes preserves a cosine-based readout. With ViT-B/16, STRA achieves 67.70%, 71.43%, and 30.43% across the three benchmark groups, exceeding the strongest listed baseline by 1.17, 1.09, and 1.87 percentage points. RN50 results support transfer across backbones.
-
Useful ablations and practical efficiency: Removing either the geometry map or prototype updates reduces accuracy. Table 24 directly compares confidence-selected and all-feature geometry estimation. Reported latency is lower than SCA and ADAPT across all three groups, with memory close to SCA and independent of stream length.
Critical weaknesses
-
Novelty needs clearer positioning: SCA’s official implementation already estimates geometry from all features independently of pseudo-label weights and uses the same cumulative-probability truncation rule. The paper should distinguish its bilateral normalization and prototype treatment from these existing components.
-
Retrofit comparisons are underspecified: Table 5 does not clarify whether spherical re-projection transforms native class prototypes or fitted classifier weights. Transforming already fitted weights can introduce an additional metric factor. Without explicit scoring formulas, the degraded retrofit results do not isolate the benefit of STRA’s estimators and readout.
Questions for authors
-
Which statistical updates are new relative to SCA? How much improvement remains when prototype blending, decay, truncation, and regularization are matched?
-
What exact scoring formulas define the Table 5 retrofits, and which class vectors are transformed?
-
How does STRA handle repeated abrupt domain changes? Does it discount or remove obsolete statistics?
-
Do you have a concrete low-rank or structured implementation, with initial results, addressing the (\mathcal{O}(d^3)) matrix decomposition cost?
-
Please reconcile two reporting discrepancies: Table 4’s unnormalized variant exceeds the frozen baseline at 66.88% versus 59.11% and 27.05% versus 25.50%, contrary to Section 4.3’s description. ADAPT’s Cars accuracy is also 69.08% in Table 2 but 68.19% in Table 26.