- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper introduces a visual explanation method for object detection, which first drives fixations from gradients and activations, then uses channel weights, local evidence, and object location gating to produce the final heatmap. The method is compared with other explanation methods over two datasets and two models, on faithfulness, localization ability, and alignment with human gaze.
Critical strengths
- This paper provides an evaluation of aligning human gaze with the explanation.
- The method is compared with several gradient-based and perturbation-based explanation methods over two datasets and two models.
Critical weaknesses
- The paper does not justify why it calculates the evidence by positive gradient * positive activation. Why not \sum ReLU(gradient*activation) or ReLU(\sum gradient * activation)?
- Selecting K fixations from the evidence inside the bounding box is not reasonable. As an explanation method, it tries to reveal how the model makes the prediction. The context region outside the bounding boxes may also contribute. Moreover, the predicted bounding box is the model’s output, which is also object attributes that requires explanation like predicted score, but not a tool to help interpret the score.
- What is the justification for using the gate to suppress distant responses while retaining nearby context? The explanation should fairly reflect the importance of each location in the image.
- Why not compare with ODAM, which I remember also proposes and provides the XAI map and human-gaze alignment results?
- The results in Table 1 haven’t shown that HiOdex has superior faithfulness and localization compared with other methods.
- The method generates highlights on the detected object by using the evidence obtained from gradients and feature activations. Although the visualized heat map looks nice, the certification that the method provides explanations for the model is weak.
Questions for authors
See the weaknesses.
13 comments
Response to Reviewer qVm2 (2/4)
W2
"Selecting K fixations from the evidence inside the bounding box is not reasonable. As an explanation method, it tries to reveal how the model makes the prediction. The context region outside the bounding boxes may also contribute. Moreover, the predicted bounding box is the model's output, which is also object attributes that requires explanation like predicted score, but not a tool to help interpret the score."
1/ "Selecting K fixations from the evidence inside the bounding box is not reasonable." We disagree; this concern comes from a misunderstanding of how the box is used. Selecting fixations inside the box is reasonable because each explanation must refer to one selected detection: a detector outputs many predictions, and each explanation must correspond to one category-and-location prediction (L51–53). Selecting fixations inside the predicted box keeps the fixation envelope, and thus the gradient pooling, confined to the selected detection (L244–245).
2/ "The context region outside the bounding boxes may also contribute." It does contribute in HiOdex; the method does not discard it. Restricting fixations to the box only determines where gradients are pooled into channel weights (Eqs. 4–5). Both branch maps are computed over the whole feature map (Eqs. 5–6), so activations outside the box still contribute.
The object-context gate is a soft Gaussian that retains nearby context (L283). The ablations show that this is better in OA than either a restrictive object mask or unrestricted spatial support (App. C.3).
Faithfulness confirms that no important context is lost. Insertion and Deletion perturb pixels across the whole image (App. B.5), yet HiOdex achieves the highest OA among the CAM-based methods and FovEx, and stays within 0.0217 OA of the strongest perturbation-based result (L351–353), even though the perturbation masks cover the entire image.
3/ Should the predicted box be explained rather than used to interpret the score? We disagree with this premise: the box is not used to interpret the score; it only identifies which prediction is explained. HiOdex explains the score of the selected prediction (L201–204).
Using the selected prediction in this way is standard for detection-specific explainers: G-CAME applies a target-centered Gaussian, and D-RISE weights its masks by their agreement with the target detection (L108–113). The box plays the same role as the class index in Grad-CAM: it selects what is explained.
W3
"What is the justification for using the gate to suppress distant responses while retaining nearby context? The explanation should fairly reflect the importance of each location in the image."
What justifies the gate? The gate is justified by two things: its design and the faithfulness results.
Design. The gate makes the explanation specific to the selected detection. It is a soft Gaussian, with a width of 0.6 of each box side (Eq. 7; Table A2), not a hard mask, so it suppresses distant responses while retaining nearby context (L283).
Faithfulness results. The gate makes the explanation reflect the importance of each location more faithfully, not less. Whether an explanation fairly reflects this importance is what OA tests: Insertion and Deletion perturb the whole image in attribution order and track the detector's score (App. B.5). Under this test, OA peaks near , removing the gate lowers OA in all four settings, and removing the gate also yields the lowest PG, EBPG and Sparsity in every setting (App. C.5, Fig. C11).
The suppressed distant responses are therefore, on average, less important to the selected detection than the ungated map suggests.
Response to Reviewer qVm2 (3/4)
W4
"Why not compare with ODAM, which I remember also proposes and provides the XAI map and human-gaze alignment results?"
Why not compare with ODAM? We disagree that a separate ODAM comparison is needed, for two reasons:
- Its mechanism is already represented. In its quantitative setting, ODAM computes for an instance-specific detection target (Zhao et al., 2024; L109–110). LayerCAM has the same position-wise gradient–activation form; the only difference is that LayerCAM rectifies the gradients (L104–105). LayerCAM is already a baseline, and in our setup it explains the same detection score as every other method (L316–317), so it is detection-specific as well.
- Its gaze results come from a small in-house study. The ODAM authors recorded eye fixations from 10 participants explaining 160 MS COCO objects, for one detector (FCOS). Our evaluation uses all eligible images of two public gaze datasets (6,202 and 6,270 images; L301–302), two detectors, and six image- and object-level metrics. This is the dataset-scale evaluation referred to in L69–70.
If the reviewer sees a specific aspect of ODAM that these baselines do not cover, we will add it to the comparison.
W5
"The results in Table 1 haven't shown that HiOdex has superior faithfulness and localization compared with other methods."
Does Table 1 show superior faithfulness and localization? We disagree with this reading of Table 1. The paper does not claim superiority on every metric; it claims the best balance of explanation quality and cost (Fig. 1), which Table 1 supports.
No method dominates every criterion, because the metrics capture different properties (L118–124). Detector explainers fall into two families with opposite trade-offs:
- CAM-based methods are fast, using at most one forward and one backward pass (L53–54), but generally less faithful (Table 1).
- Perturbation-based methods can provide strong faithfulness, but need hundreds or thousands of detector evaluations per explanation (L114–116).
HiOdex combines the speed of the first family with faithfulness close to the second:
- Faithfulness. HiOdex achieves the highest OA among the CAM-based methods and FovEx in all four settings, and stays within 0.0217 OA of the strongest perturbation-based result (L351–353). It is approximately two orders of magnitude faster than D-RISE and D-CLOSE (L367–369), and no evaluated method achieves both a lower generation time and a higher OA (Fig. 1). It also ranks among the top three on both Deletion and Insertion in all four settings (App. C.1).
- Localization. HiOdex achieves the highest PG with Faster R-CNN on both benchmarks. With YOLOv12-N, the leading PG values exceed its scores by only 0.24 and 1.45 percentage points. It ranks third on EBPG in all four settings; the methods ahead of it produce more concentrated maps, and concentrated attribution is not sufficient for faithfulness (App. C.1).
- Overall. HiOdex ranks among the top three in 22 of 24 comparisons and never below fourth (L353–354).
- Human gaze. HiOdex achieves the best or tied-best AUC-Judd in all four gaze settings (Table 2), and the best average rank across AUC-Judd, SIM and KL divergence (L428–429).
Response to Reviewer qVm2 (1/4)
We thank the reviewer for acknowledging our human-gaze evaluation and the comparison with gradient- and perturbation-based methods across two datasets and two detectors.
However, we strongly disagree with this review. Its main criticisms are contradicted by the paper itself: they rest on misreadings of the method and of the paper's claims, and they overlook the quantitative evidence the paper reports. None of the weaknesses identifies an error in the method or the results. The clear-rejection rating and the soundness score of 1 are therefore unfounded. We ask the reviewer to re-examine each point against the evidence cited below and to correct the assessment.
Regarding the weaknesses and questions raised, please find our point-by-point responses below.
W1
1/ Does the paper justify it? Not explicitly, so we justify it here: Eq. 1 is the only one of the three forms that satisfies the two conditions evidence must meet. The evidence map decides where fixations are placed (L224–226), so a location should count as evidence for a channel only when:
Eq. 1 keeps exactly the channel–location pairs that satisfy both conditions.
2/ Why not ∑ReLU(G⋅A)? Because it violates condition 1. It equals Eq. 1 wherever A≥0, but where A<0 and G<0 the product is positive. An absent or suppressed feature is then scored as evidence, so fixations could land where nothing supports the object.
3/ Why not ReLU(∑G⋅A)? Because it does not check the two conditions per channel at all. It sums signed products across channels before rectifying, so opposing channels can cancel strong supporting evidence at the same location, while negative–negative pairs still count. It measures a net effect, not where supporting evidence lies.
Signed channel aggregation is already used where it belongs: the Foveal CAM computes ReLU(∑kαkAk) (Eq. 5). The Local Evidence Map follows the same two conditions as Eq. 1 (Eq. 6).