- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper addresses test-time adaptation (TTA) for vision-language object detectors (VLODs) within a cumulative mean-teacher framework. The authors identify two primary failure modes in standard mean-teacher OD-TTA: extreme domain sensitivity to a fixed exponential moving average (EMA) teacher update coefficient, and long-term student parameter drift caused by repeated optimization on noisy pseudo-labels. To resolve these issues, the authors propose DESA-TTA, which combines two heuristics: (1) Dynamic Temporal Averaging, which computes a sample-wise teacher uncertainty score from mean pseudo-label confidence and normalized box density to interpolate the EMA coefficient within bounds controlled by teacher parameter drift; and (2) Source Anchoring, which linearly interpolates gradient-updated student normalization parameters back toward their pretrained initialization based on an student parameter drift metric. The method is validated across two VLOD architectures (YOLO-World and Grounding DINO) on style shifts (Watercolor, ClipArt, Comic), driving scenes (Foggy Cityscapes, BDD100K), low-light (ExDark), and common corruptions (VOC-C, COCO-C), demonstrating consistent and mAP improvements alongside favorable inference throughput.
Critical strengths
- Practical Efficiency and Empirical Gains in VLOD-TTA: Adapting open-vocabulary object detectors online is notoriously brittle and computationally expensive. By restricting updates to normalization layers and regulating the teacher-student trajectory via closed-form parameter interpolation and dynamic EMA scaling, DESA-TTA achieves substantial accuracy gains (+8.9 over VLOD-TTA on VOC-C) while increasing FPS from 20 to 31 on YOLO-World and cutting memory consumption by 70.4% compared to foundation-model-assisted baselines like TTAOD-F on Grounding DINO.
- Thorough Empirical Diagnostics: The paper does an admirable job of empirically motivating its design choices. The image-level Spearman correlation analysis (Table 7) and quintile analysis (Table 8) verifying that the proposed heuristic uncertainty score negatively correlates with true pseudo-label precision and IoU—even when controlling for ground-truth object count—provides solid empirical grounding for an otherwise heuristic rule.
Critical weaknesses
- Heavy Parameterization of Update Heuristics and Validation Leakage Concerns: While the high-level intuition of DESA-TTA is clean, its actual instantiation introduces 9 distinct hyperparameters (, and ) governed by ad-hoc rational transition functions (Equations 4–9). The authors state these were tuned on "six corruption types applied to the VOC training split" and kept fixed. However, tuning on synthetic corruptions of PASCAL VOC and then claiming strong performance on VOC-C (which shares the exact same underlying source imagery and object statistics, plus overlapping corruption families) represents a favorable domain overlap. Moreover, Figure 5 reveals that optimal hyperparameters diverge between short style-shift streams and long corruption streams (e.g., and have opposing trends after a threshold), suggesting the method's drift scales are implicitly coupled to stream length.
- Counter-Intuitive Uncertainty-to-EMA Logic Under Severe Corruptions: The core premise of Dynamic Temporal Averaging is that high teacher uncertainty (, meaning dense, low-confidence boxes) should yield a smaller EMA coefficient , thereby incorporating the current student state more rapidly. However, the student was just updated using those exact same dense, low-confidence, noisy pseudo-labels from sample . Accelerating the injection of a student step supervised by highly uncertain pseudo-labels seems risky. In Proposition 1 (Appendix A.1), this is only justified by assuming the student error variance remains fixed while teacher error variance grows—an assumption that breaks down in practice since directly depends on the quality of the teacher's supervision at step .
Questions for authors
- Clarification on Student-Teacher Error Coupling: In Proposition 1, you assume and are fixed with respect to . In practice, when teacher uncertainty spikes on a heavily corrupted frame, the student gradient step is computed directly from those degraded pseudo-labels. Why does lowering (faster student-to-teacher injection) on high- samples not accelerate teacher corruption? Is this solely because the student has just been pulled back toward via Source Anchoring?
- Clarification on Grounding DINO Hyperparameters: Were the exact same hyperparameters (such as ) tuned on YOLO-World (BatchNorm) directly transferred without modification to Grounding DINO (LayerNorm/GroupNorm) in Table 3? Given that and measure relative parameter drift from initialization, do affine normalization weights in Transformers drift at the same numerical scale as CNN BatchNorm statistics?
- Suggestion for Future Work / Continual TTA: Since the drift scales () are calibrated around a specific stream horizon, how would DESA-TTA behave in a continual TTA setting (e.g., sequenced corruptions without resetting to )? If saturates in Domain 1, will be locked near , potentially preventing adaptation to Domain 2.
5 comments
Response to Reviewer 7EKG (2/3)
W2 / Q1. Uncertainty-to-EMA logic and student–teacher coupling.
We agree that the student update on uses the same pseudo-labels whose uncertainty determines , so and may vary with . The expression for the unconstrained one-step optimum in Proposition 1 remains valid under this dependence, which is explicitly included through . Writing and , we obtain
For and , this coefficient decreases with whenever increases. The student error therefore need not remain constant. What matters is how the ratio of these conditional error terms changes with uncertainty. Holding and fixed, with and increasing , is one sufficient special case. We have stated this general condition and clarified the conditional nature of the analysis in Proposition 1 (Appendix A.1) of the revised PDF.
Control of the student contribution.
Updating the teacher from a student trained on the teacher's own pseudo-labels is the standard mean-teacher mechanism. Because the teacher averages many past student states, it provides more stable supervision than the online student. This mechanism is widely used in semi-supervised object detection [4] and test-time adaptation for object detection [2]. In our experiments, MT improves over ZS on five of the six benchmarks in Table 1 and on VOC-C (Table 2). DESA-TTA retains this averaging and modulates the update rate only within bounded limits:
- Bounded contribution. is at least . Each teacher update therefore assigns at most 0.7% weight to the current student state.
- Drift-dependent range. As teacher drift increases, both EMA bounds rise toward . The upper bound on the student contribution consequently approaches 0.1% in the high-drift limit.
- Source anchoring. After each gradient step, the student is interpolated toward its source initialization before the teacher update. This contracts its post-gradient parameter distance from the source (Proposition 2).
Additional Benefit of Dynamic EMA Beyond Source Anchoring
With source anchoring disabled, D-EMA improves over fixed EMA:
| Method | Style AP50 | VOC-C AP50 |
|---|---|---|
| Fixed EMA (MT, Table 5) | 44.1 | 39.5 |
| Per-domain fixed-EMA oracle | 45.4 | 42.5 |
| D-EMA (Table 5) | 46.6 | 43.1 |
D-EMA improves over MT by 2.5 AP50 on Style and 3.6 AP50 on VOC-C, and also exceeds the per-domain fixed-EMA oracle from our additional sweep. These comparisons show that the improvement does not depend solely on restoring the student toward the source.
Empirical test of the opposite rule.
Table 9 tests the reversed mapping, which increases the EMA coefficient on uncertain samples, using the same dynamic source-anchoring rule and drift-dependent EMA-bound functions. It performs worse than DESA-TTA on all 18 datasets and corruptions, by 2.6 AP50 on Style and 5.2 AP50 on VOC-C on average. This directly supports the proposed mapping direction under otherwise matched settings. Tables 7 and 8 further support the use of as an indicator of pseudo-label quality.
Q2. Hyperparameters for Grounding DINO and drift scale.
YOLO-World and Grounding DINO are different architectures: YOLO-World is a CNN with batch normalization, whereas Grounding DINO is a Transformer with layer and group normalization. As the reviewer notes, their normalization parameters do not necessarily drift at the same numerical scale. For Grounding DINO, we use a learning rate of and set , , , , , , , , , and . With these values, DESA-TTA improves the average COCO-C mAP from 20.6 to 27.6 and outperforms TTAOD-F and VLOD-TTA by 1.6 and 1.4 mAP, respectively (Table 3). We have added these values to Appendix A.2 of the revised PDF.
Response to Reviewer 7EKG (3/3)
Q3. Continual TTA and drift saturation.
We thank the reviewer for this suggestion. Our method focuses on single-domain TTA, following prior VLOD TTA work, VLOD-TTA [1] and TTAOD-F [2]. Continual adaptation across sequential domains is outside the scope of our main evaluation.
As teacher drift increases, the EMA bounds approach . In this regime, the teacher continues to update, with an approximate EMA memory scale of 1,000–2,500 updates, comparable to the best fixed EMA on VOC-C (). The student's gradient update is not directly scaled by , although it remains coupled to the teacher through the pseudo-labels.
After an abrupt shift, however, this slower teacher response could delay adaptation to a new domain, because is measured relative to the source model and does not necessarily decrease when the new domain begins.
To examine this, we evaluated DESA-TTA on the 15 VOC-C corruptions as a single continual stream, without resetting between corruptions. We also combined it with a shift detector based on the feature-statistics criterion of CTTAOD [3] (), which resets the student and teacher to the source model when a shift is detected. Since DESA-TTA does not access source data, the training-feature statistics required by this criterion are taken from the batch-normalization running statistics stored in the pretrained YOLO-World model. These results are reported in Appendix A.3 (Sequential Domain Shifts, Table 15) of the revised PDF. The last column lists the image index within each corruption at which a reset occurred.
| Corruption | Single-domain (Table 2) | Continual, no reset | Continual, shift detection | Resets (image index) |
|---|---|---|---|---|
| Gaussian noise | 34.4 | 34.4 | 34.4 | – |
| Shot noise | 35.8 | 38.2 | 38.2 | – |
| Impulse noise | 37.8 | 38.8 | 38.8 | – |
| Defocus blur | 39.6 | 34.4 | 39.2 | 11 |
| Glass blur | 32.7 | 34.9 | 34.1 | – |
| Motion blur | 37.8 | 35.9 | 36.8 | – |
| Zoom blur | 36.1 | 36.3 | 38.0 | – |
| Snow | 54.5 | 51.9 | 52.4 | – |
| Frost | 57.8 | 58.6 | 57.7 | 2165, 4734, 5234 |
| Fog | 73.6 | 72.3 | 72.8 | 74 |
| Brightness | 74.2 | 73.9 | 74.2 | 41, 895, 1530, 3583, 4083 |
| Contrast | 63.3 | 62.0 | 62.2 | 8 |
| Elastic transform | 59.8 | 56.3 | 55.6 | 17, 4514, 5210 |
| Pixelate | 37.9 | 17.6 | 33.1 | 14 |
| JPEG compression | 52.6 | 52.1 | 53.0 | – |
| Average | 48.5 | 46.5 | 48.0 | 15 resets |
Without resetting, DESA-TTA achieves 46.5 AP50, compared with 48.5 under the single-domain protocol, and remains well above MT (39.5) and VLOD-TTA (39.6) evaluated in the single-domain protocol. It exceeds its single-domain results on five corruptions, consistent with beneficial carry-over across some domains. However, the effect varies across corruptions, with Pixelate decreasing from 37.9 to 17.6 AP50. With shift detection, the average improves to 48.0 AP50, within 0.5 of the single-domain result, and Pixelate recovers to 33.1 AP50. The detector triggers within the first 75 images of 6 of the 14 new corruptions, with 9 additional resets occurring later within corruptions. These results show that source resets triggered by shift detection mitigate the performance loss in the tested continual sequence, and that more accurate shift detection is a natural direction for future work.
References
[1] A. Belal, H. R. Medeiros, M. Pedersoli, and E. Granger. VLOD-TTA: Test-time adaptation of vision-language object detectors. In ECCV, 2026.
[2] Y. Gao, Y. Zhang, Z. Cai, and D. Huang. Test-time adaptive object detection with foundation model. In NeurIPS, 2025.
[3] J. Yoo, D. Lee, I. Chung, D. Kim, and N. Kwak. What, how, and when should object detectors update in continually changing test domains? In CVPR, 2024.
[4] Y.-C. Liu, C.-Y. Ma, Z. He, C.-W. Kuo, K. Chen, P. Zhang, B. Wu, Z. Kira, and P. Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2021.
Response to Reviewer 7EKG (1/3)
Thank you for your detailed and constructive review, and for recognizing the practical efficiency of DESA-TTA and the empirical diagnostics of the teacher-uncertainty score. Your questions about the hyperparameters, the tuning protocol, and the coupling between student and teacher errors helped us clarify the assumptions behind our design. We particularly appreciate your suggestion to study continual adaptation, which led to the new continual VOC-C experiment with shift detection below. Our modifications are highlighted in blue in the revised PDF.
W1. Hyperparameters, tuning protocol, and stream length.
Number of tuned hyperparameters.
We tune five hyperparameters, and the remaining quantities are fixed across all target domains:
All drift-dependent terms use the same monotone saturating map, δp/(δp+λp). Each scale λ has a direct interpretation: it is the drift at which the corresponding transition is half complete. A single exponent p is shared by all three transitions.
Sensitivity.
Fig. 5 varies each tuned hyperparameter independently while keeping the others fixed. Across these sweeps, AP50 stays within approximately 1 point of the selected configuration on both benchmarks (VOC-C 47.5–48.6, Style 46.3–47.5). Every tested setting remains above MT (39.5 VOC-C, 44.1 Style) and VLOD-TTA (39.6 VOC-C, 44.9 Style). These ranges are now also stated in Appendix A.4 of the revised PDF.
Tuning data and overlap with VOC-C.
The hyperparameters were selected on six corruption types applied to the VOC training split, which shares no images with the VOC-C evaluation set. We acknowledge that the tuning and evaluation data share object categories and corruption families, so VOC-C results alone cannot establish generalization. However, the same configuration is also evaluated unchanged on six benchmarks with no image overlap with the tuning data. They cover artistic (Watercolor, ClipArt, Comic), adverse-weather (Foggy Cityscapes), real-world driving (BDD100K), and low-light (ExDark) shifts. DESA-TTA achieves the best AP50 on all six, improving over VLOD-TTA by +1.9, +2.9, +2.5, +2.2, +3.5, and +1.8 AP50, respectively (Table 1).
Stream length.
We agree that the best values of αmax and λmin differ between Style and VOC-C (Fig. 5a, c). However, the two benchmarks differ in both the type of shift and the stream length, so these differences cannot be attributed to stream length alone. Moreover, DESA-TTA sets source anchoring and the EMA bounds from the measured student and teacher drift rather than from the number of processed images.
Furthermore, a single configuration is effective for streams ranging from 500 images (ClipArt, Foggy Cityscapes) to 10,000 images (BDD100K). On the two longest streams, BDD100K and ExDark, neither of which was used for tuning, DESA-TTA improves over MT by +7.5 and +3.8 AP50. On BDD100K, MT itself falls below ZS.