- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
PAVER targets tiny-object detection in remote-sensing imagery. Its detector modifies the feature pathway so that weak spatial cues are less likely to disappear during deeper processing. A later confidence stage reads P2/P3 ROI features and uses them to alter candidate scores.
CSFM, SCRG, and HRPE form the detector. The Risk-conditioned Interface is fitted after detector training. Its prototype margin is combined with the detector score, while the detected boxes are left unchanged.
Detection is evaluated on VisDrone, RSOD, and DIOR. VisDrone gives the largest gain: 45.9 mAP50 for PAVER versus 40.2 for YOLOv12-s, with mAP50:95 increasing from 24.0 to 28.0. Confidence quality is studied separately on VisDrone test-dev.
The detector experiments are the more convincing part of the paper. The confidence results are positive relative to raw scores, although the comparison with ordinary score calibration is much less clear.
Critical strengths
The VisDrone detection result is strong. The gain remains visible under stricter localization rather than appearing only at IoU 0.50. AP75 increases from 25.8 to 31.4. APS rises from 16.3 to 21.6.
The architectural study is also fairly extensive. HRPE provides the largest single improvement. The paper then compares the proposed backbone replacement against other blocks, and performs similar controlled comparisons for the fusion and pyramid designs.
These experiments are useful because the full detector modifies several locations at once. The matched replacements make it easier to see whether a gain comes only from adding capacity or from the selected architectural choice.
Detection and confidence calibration are evaluated separately. The confidence study begins from fixed detector outputs, so its results are not mixed with the detector's AP gain.
The fitting protocol is also careful in several respects. Calibration uses the validation set. Test-dev annotations are used for evaluation rather than parameter fitting. Bootstrap resampling is carried out at the image level.
Critical weaknesses
The main concern is the evidence for the Risk-conditioned Interface. PAVER improves the calibration point estimates relative to raw detector confidence, but the more informative comparison is with Platt scaling, which is fitted on the same validation labels and is nested in the proposed formulation when the risk coefficient is zero. Against Platt, the reported differences are only about -0.00015 for ECE, -0.00003 for Brier score, and -0.00014 for NLL, and all three paired 95% intervals include zero. Temperature scaling also achieves lower ECE and NLL than PAVER. PAVER has the best Brier point estimate and FP-AUC, but its FP-AUC of 0.94466 differs from the score-only alternatives at 0.94456 by only , without an analogous paired interval. The current results therefore establish an improvement over raw confidence more clearly than an additional benefit from ROI evidence beyond standard score calibration.
A second concern is the definition of candidate correctness. TP/FP labels are generated by greedy matching in raw-score order and then kept fixed across calibration variants, although PAVER can change the candidate ranking. For duplicate detections, a different ordering can change which candidate claims an available ground truth, so a candidate labelled FP before recalibration need not receive the same label if matching is repeated afterward. The calibration metrics consequently evaluate confidence against a correctness event defined under the original ranking, whereas AP is recomputed after the scores change. This protocol is reasonable if that fixed event is the intended target, but it is harder to interpret as correctness of the final ranked output. Repeating TP/FP matching after each recalibrated ranking, and reporting how often assignments change, would clarify this issue.
The confidence study is also narrow in both dataset coverage and uncertainty estimation. The Risk-conditioned Interface is evaluated only on VisDrone, although detection experiments are also reported on RSOD and DIOR. A prototype relation learned from P2/P3 ROI features may depend on object scale, background statistics, or acquisition characteristics, so cross-dataset calibration would be important for a component framed in terms of confidence trustworthiness. In addition, the bootstrap resamples test-dev images while holding the fitted normalizer, prototype banks, and calibration parameters fixed. The intervals therefore describe test-image sampling uncertainty conditional on one fitted interface and omit variation from refitting the confidence model. This omission is relevant because the PAVER-versus-Platt differences are already close to zero. The validation protocol is image-disjoint, but scene-disjointness is not established because suitable scene-group metadata are unavailable.
The evidence-preservation interpretation also needs more direct support. Tables 5–7 show that the proposed architectural components improve AP and that HRPE in particular contributes a sizeable gain. These experiments establish the utility of the design, but they do not directly measure whether weak-target information is preserved through the feature pathway. Figure 1 provides qualitative activation examples, yet no quantitative feature-level diagnostic is reported. Measuring target/background separability across stages, or tracking the retention of small-object responses before and after the relevant transformations, would provide a more direct test of the mechanism. Without such evidence, “evidence preservation” remains a plausible explanation for the AP improvement rather than a verified property of the learned representation.
Efficiency is another concern. At , PAVER requires 40.0 GFLOPs compared with 21.4 for YOLOv12-s, while detector-core latency on the reported T4 increases from 2.61 to 4.89 ms per image and peak memory from 0.82 to 1.31 GB. These measurements exclude the Risk-conditioned Interface and the operations required by it, so the end-to-end cost of the complete proposed system is not reported. The comparison with YOLOv12-s at 1024 provides useful context because PAVER at 640 obtains slightly better accuracy at lower nominal GFLOPs, but complete latency remains necessary for evaluating the practical cost of the full pipeline.
The cross-dataset detection evidence is also mixed. PAVER performs strongly on RSOD, reaching 96.7 mAP50 and 68.9 mAP50:95, but on DIOR it matches YOLOv12-s at 74.5 mAP50 while trailing it at mAP50:95, 53.0 versus 55.0. The DIOR evaluation partition also participates in checkpoint selection, so this result is better treated as a controlled internal comparison than as a blind external benchmark. Some matched architectural replacements in Table 6 differ by only 0.2–0.3 AP, yet the paper reports a common seed rather than repeated-run variability for these close comparisons. This is less problematic for the large HRPE gain, but it makes the ordering among closely performing alternatives harder to judge.
Overall, the two main contributions are supported at different levels. The detector has sizeable VisDrone gains, stricter-localization improvements, and reasonably extensive component studies. The Risk-conditioned Interface has favorable point estimates, but its incremental advantage over Platt scaling remains statistically unresolved and is evaluated on only one dataset. Because the confidence interface receives substantial emphasis in the title, abstract, and overall framing, this gap affects the strength of the main contribution rather than only an auxiliary experiment.
Questions for authors
Can a paired 95% interval be reported for FP-AUC when comparing PAVER with Platt scaling?
Can the PAVER-versus-Platt bootstrap be repeated while refitting the calibration stage inside each replicate?
How many TP/FP assignments change after PAVER reorders candidate scores?
What happens to ECE, Brier, and NLL if greedy matching is repeated using each method's final score ordering?
Can the Risk-conditioned Interface be evaluated on RSOD or DIOR?
Is there a quantitative feature-level test for the proposed evidence-preservation mechanism?
For the close replacements in Table 6, does the ranking remain the same across independent training runs?
What is the end-to-end latency after the Risk-conditioned Interface is included?
How sensitive are the calibration results to prototype budget ?
Do P4 or P5 features give a similar calibration result to P2/P3?