acceptodds
Withdrawn from ICLR 2027

A Gaze-Free Human-Inspired Object Detection Explainer via Evidence-Based Fixations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Explanations for object detectors are usually evaluated against detector behavior, leaving their alignment with human attention unclear. Human-inspired explainers largely target image classification, while human-supervised methods learn from gaze and cannot use it as an independent reference. We introduce HiOdex (Human-inspired Object Detection Explainer), the first post-hoc, human-inspired explainer that produces object-specific attributions for object detectors without gaze supervision. HiOdex derives model-generated fixations from activation- gradient evidence, applies inhibition of return, and uses their envelope to guide class activation mapping (CAM). Local evidence and multi-layer fusion preserve spatial detail in a single forward and backward pass. We compare HiOdex with eight explainers across two detectors and two benchmarks using model-oriented metrics, then independently assess human-gaze alignment on two gaze datasets using image-level saliency metrics and object-level measures. HiOdex achieves the highest Over-All score among all evaluated CAM-based and human-inspired baselines in all four detector and benchmark settings. It comes within 0.0217 of the strongest perturbation-based result at roughly 1% of the computational cost, achieves the best or tied-best AUC-Judd in every gaze setting, and ranks in the top two for 8 of 12 evaluations using image-level gaze metrics. These results show detector-derived fixations can yield efficient explanations that perform strongly under both model- and human-centered evaluation without gaze supervision.

Primary area:
interpretability and explainable AI
Keywords:
Explainable artificial intelligence, object detection, human-inspired explanations, human gaze alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qVm2OpenReview ↗
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 (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

"The paper does not justify why it calculates the evidence by positive gradient * positive activation. Why not ∑ReLU(G⋅A)\sum \mathrm{ReLU}(G \cdot A) or ReLU(∑G⋅A)\mathrm{ReLU}(\sum G \cdot A)?"

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:

  1. the feature is present (A>0A > 0);
  2. increasing it would raise the detection score sτs_{\tau} (G>0G > 0).

Eq. 1 keeps exactly the channel–location pairs that satisfy both conditions.

2/ Why not ∑ReLU(G⋅A)\sum \mathrm{ReLU}(G \cdot A)? Because it violates condition 1. It equals Eq. 1 wherever A≥0A \geq 0, but where A<0A < 0 and G<0G < 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)\mathrm{ReLU}(\sum G \cdot 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)\mathrm{ReLU}(\sum_{k} \alpha_{k} A_{k}) (Eq. 5). The Local Evidence Map follows the same two conditions as Eq. 1 (Eq. 6).

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 sτs_{\tau} of the selected prediction τ=(bτ,cτ,sτ)\tau = (b_{\tau}, c_{\tau}, s_{\tau}) (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 ρ=0.6\rho = 0.6, 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 ReLU(∑k∂s/∂Ak⊙Ak)\mathrm{ReLU}(\sum_{k} \partial s / \partial A_{k} \odot A_{k}) 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 sτs_{\tau} 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).
Reviewer_GSJDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces an attribution method for individual object detection named HiOdex. This is connected to measurements of saliency in human vision, and positive activation gradients facilitate selection of a short sequence of feature space fixations with an inhibition of return approach discouraging repeated fixations. Spatial pooling is performed into CAM channels via mixture of Gaussians, and fused across layers. Eight baselines are compared against towards establishing the value of the attribution approach assessed via human gaze and segmentation type datasets..

Critical strengths
  1. This work provides a useful approach for fast attribution with the possibility of using cached activations / gradients - This is a positive, and could see use of the model on the basis of its apparent simplicity and reasonably good performance shown
  2. The model provides improvements over CAM based methods tested for this particular type of assessment, and might be of value where explainability tied to gaze or objects is desired
  3. The assessment is broad in including several domains, and two independent gaze datasets which lends confidence to the results. There is promise shown in the scores associated with all of the benchmarks performed, and this suggests value from the work that is presented in the paper.
Critical weaknesses
  1. The human alignment claim is tied to the fixation process, and fixations and inhibition of return are an important mechanism of the work, but it appears that some of what is presented suggests local evidence on its own (e.g. local feature activations) is competitive. It may be relatively important to provide clarity on this since it relates to a central claim and no ablations seem to address human gaze that considers the role of features as opposed to the gaze mechanism.
  2. It would be helpful to know what the inclusion/exclusion criterion for attribution methods that were considered look like. There is clearly a structural relationship in these approaches to that presented in the paper, but it would be interesting to know what this means broadly for attribution for the metrics considered across all methods. For example, should these be compared with methods from the human gaze prediction literature? Should they be compared with objectness measures, or literature around neural network attribution across other vision tasks in considering methods like attention rollout, or Chefer's work. It's not clear what the appropriate comparison is, but the authors should make this case.
  3. I found this paper hard to follow in places since it mixes many different concepts. This is evident from the beginning - e.g. "Explanations for object detectors are usually evaluated against detector behavior, leaving their alignment with human attention unclear. Human-inspired explainers largely target image classification, while human-supervised methods learn from gaze and cannot use it as an independent reference." - the first says that object detectors are evaluated against detector behavior which makes sense, and then goes on to say that the alignment with human attention is unclear, which is fine too. But then the jump to "human inspired explainers" targeting image classification, and human supervised methods learning from gaze. There are many different ideas mixed here that makes it hard to track what is being said (for me at least).
Questions for authors
  1. How important do you feel the connection of the current work is to human gaze behaviour as compared to object detection behaviour in a typical attribution sense? For example, what do we learn that attribution aligned with human gaze says as compared to segmentation or perturbation measures?
  2. How strong is local evidence in predicting gaze in comparison to the proposed attribution approach, and what is learned about the proposal with respect to features the model uses in decision making, and those that associate with human gaze data (e.g. are local gradients at different spatial frequencies enough, and what does the gaze mechanism itself bring)?
  3. In figure 2, it feels like the regions have a horizontal stretch to them - is this something related to spatial bias of the attention mechanism?
5 comments

Response to Reviewer GSJD (1/4)

We thank the reviewer for recognizing HiOdex's fast attribution from cached activations and gradients, its improvements over CAM-based methods, and the breadth of our evaluation across domains and two independent gaze datasets.

Regarding the weaknesses and questions raised, please find our point-by-point responses below.

W1

"The human alignment claim is tied to the fixation process, and fixations and inhibition of return are an important mechanism of the work, but it appears that some of what is presented suggests local evidence on its own (e.g. local feature activations) is competitive. It may be relatively important to provide clarity on this since it relates to a central claim and no ablations seem to address human gaze that considers the role of features as opposed to the gaze mechanism."

1/ Does the competitiveness of local evidence weaken the human-alignment claim? We disagree that it does. Local evidence is competitive only on the model-oriented metrics, which cannot assess human alignment:

  • The claim concerns the full HiOdex. The paper attributes HiOdex's gaze agreement to detector-derived evidence together with the spatial organization imposed by the fixation process (L453–455), not to the fixation process alone. The Local Evidence Map is part of that evidence: it is one of HiOdex's two branches (L260–262).
  • The quantitative metrics alone cannot separate the two branches. The ablations report only model-oriented metrics and do not measure gaze alignment (App. C.2). As the paper notes, these metrics do not establish whether an intermediate mixture outperforms the Local Evidence Map alone (App. C.5). The Local Evidence Map's higher EBPG and Sparsity (Fig. C9) reflect more concentrated maps, and a concentrated map is not necessarily correct (App. B.5).
  • The qualitative maps show the difference (Fig. C10). With the Local Evidence Map alone (β=0\beta = 0), attribution forms one dominant region, such as the lower horse's torso, or the person's lower face, neck and collar. With β=0.4\beta = 0.4–0.50.5, the Foveal CAM extends attribution to other evidence-bearing parts, such as the horse's head, and centers it on the person's face. This is the transition from localized evidence to the broader Foveal CAM response (App. C.5), and it is why the paper concludes that fixation-guided context should complement, not replace, local evidence (App. C.3).

2/ Is there a gaze ablation of features versus the gaze mechanism? We agree that it is missing. We will add to the appendix a gaze-metric comparison, following the Table 2 protocol, of full HiOdex with two variants:

  • the Local Evidence Map alone (β=0\beta = 0), which removes the fixation process;
  • a single fixation (K=1K = 1), which removes inhibition of return.

Response to Reviewer GSJD (2/4)

W2

"It would be helpful to know what the inclusion/exclusion criterion for attribution methods that were considered look like. There is clearly a structural relationship in these approaches to that presented in the paper, but it would be interesting to know what this means broadly for attribution for the metrics considered across all methods. For example, should these be compared with methods from the human gaze prediction literature? Should they be compared with objectness measures, or literature around neural network attribution across other vision tasks in considering methods like attention rollout, or Chefer's work. It's not clear what the appropriate comparison is, but the authors should make this case."

We clarify the criteria, what the structural relationship means for the metrics, and why the suggested families do not fit.

1/ Inclusion/exclusion criteria. They follow from one requirement: gaze must remain an independent reference. A method is included if it is post-hoc, produces an attribution map for a selected detection, and uses no human gaze. This gives three families (L84–85): CAM-based (Grad-CAM, Grad-CAM++, LayerCAM, Eigen-CAM, G-CAME), perturbation-based (D-RISE, D-CLOSE) and human-inspired (FovEx) (L310–313).

Human-supervised methods such as HAG-XAI and HINT are excluded, because the attention data they are trained on cannot serve as an independent evaluation reference (L64–68).

2/ What the structural relationship means for the metrics. The conclusions do not depend on it, because the baselines are not limited to structurally related methods. Only the CAM-based baselines are structurally related to HiOdex: D-RISE and D-CLOSE infer importance from changes in detector output (L111–112), and FovEx iteratively foveates the input (L148–150).

All nine methods are scored with the same metrics on the same detections (L316–317). Across these different mechanisms, no method dominates every metric (L123–124), and HiOdex ranks among the top three in 22 of 24 comparisons (L353–354).

3/ Gaze-prediction models? No, they are not suitable baselines. They predict where people look, not why the detector made a detection. They are also typically trained on fixations, which would remove the independence of the gaze reference.

4/ Objectness measures? No, they are not suitable baselines. They indicate where objects are, independent of the detector's decision for a selected detection.

5/ Attention rollout or Chefer et al.'s work? No, they cannot be applied to both detectors. These methods are designed for attention-based (transformer) models, whereas Faster R-CNN uses a ResNet-50 FPN backbone (L297–298).

Response to Reviewer GSJD (3/4)

W3

"I found this paper hard to follow in places since it mixes many different concepts. This is evident from the beginning - e.g. "Explanations for object detectors are usually evaluated against detector behavior, leaving their alignment with human attention unclear. Human-inspired explainers largely target image classification, while human-supervised methods learn from gaze and cannot use it as an independent reference." - the first says that object detectors are evaluated against detector behavior which makes sense, and then goes on to say that the alignment with human attention is unclear, which is fine too. But then the jump to "human inspired explainers" targeting image classification, and human supervised methods learning from gaze. There are many different ideas mixed here that makes it hard to track what is being said (for me at least)."

The two sentences form a single argument; we agree that the link between them is implicit, and we will state it explicitly in the abstract. Each part means the following:

  • "Explanations for object detectors are usually evaluated against detector behavior": existing metrics, such as Insertion, Deletion and the Pointing Game, compare an explanation with detector behavior or object annotations (L56–57, L118–126).
  • "leaving their alignment with human attention unclear": these metrics cannot show whether the explanation highlights the regions a person would inspect (L49–51).
  • "Human-inspired explainers largely target image classification": these methods reproduce human perceptual mechanisms, such as fixations, without learning from recorded gaze (L147–148). The existing one, FovEx, primarily addresses image classification, with only a limited object-detection study (L62–64).
  • "human-supervised methods learn from gaze and cannot use it as an independent reference": methods such as HAG-XAI and HINT are trained on recorded human attention, so their agreement with gaze is no longer independent evidence (L64–68).

The implicit link is that neither group offers a gaze-free explainer for object detection, so alignment with human gaze cannot yet be tested independently. HiOdex fills this gap (L68–71).

Q1

"How important do you feel the connection of the current work is to human gaze behaviour as compared to object detection behaviour in a typical attribution sense? For example, what do we learn that attribution aligned with human gaze says as compared to segmentation or perturbation measures?"

1/ How important is gaze compared with detector behavior? Equally important, because they answer different questions. Perturbation and localization measures (Insertion, Deletion, OA, PG, EBPG) test whether the highlighted regions matter to the detector and lie on the object. They do not test whether the explanation highlights the regions a person would inspect (L49–51).

2/ What do we learn from gaze-aligned attribution compared with segmentation or perturbation measures? Two things that those measures cannot show:

  • an independently collected human reference (L58–59);
  • structure within the object: OFCS and WDSIM test whether attribution covers human-consensus locations, and in proportions consistent with human consensus (App. B.6).

The two views can also rank methods differently. For example, D-RISE has the best OA with YOLOv12-N on MS-COCO (Table 1), but ranks eighth in SIM on COCO-Search18 (Table 2). Strong model-oriented scores therefore do not imply human alignment, which is why the paper reports both.

Reviewer_pZUrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an XAI method for object detection tasks which takes into account human-inspired metrics based on human gaze. While gaze is traditionally not taken into account like in grad-CAM like methods, the proposed one takes into account a basic scanpath computed on the ROI of the object map along with other improvements.

Critical strengths

The main strength is in the idea. Taking predicted human-like gaze into account is definitely a good path towards better generic image XAI.

Critical weaknesses

The validation is made on reasonable datasets and metrics given the domains and results are state of the art but not the best. I would also expect some critique on the different metrics which are maybe not so suited : how to compare in a fait way CAM maps is not obvious.

Questions for authors

clarification requests : It is important to well define pretty much everything as your reader can come from different communities :

line 73 not obvious for people not used to cam-like maps. For the maps we found attribution, the attribution maps without any definition of the attribution map. If this is obvious for someone working in XAI it is not for someone working in saliency for example.

73 a differentiable detector is defined later at line 201, again a term which should be briefly defined at first use. Again, obvious for XAI but not necessarily for others.

120 I would add : "gradually" (or equivalent) revealed or removed. Also revealed and removed given what ? Given the importance in the map ?

133 which and not whose if you refer to yolo ...

work : Definitely what is needed is :

  • a discussion on the metrics stability (given different filtering techniques on the different maps for example)
  • there are plenty of ways to extract potential fixations from saliency maps and several approaches might be taken into account
  • the validation is convincing but there is no "wow"
2 comments

Response to Reviewer pZUr (1/2)

We thank the reviewer for the positive assessment, and for finding the idea promising and the validation convincing.

Regarding the weaknesses and questions raised, please find our point-by-point responses below.

W1

"The validation is made on reasonable datasets and metrics given the domains and results are state of the art but not the best."

"…state of the art but not the best." Not being the best on every metric is not a weakness here: no evaluated method is the best on every metric, because the metrics capture different properties (L118–124). The paper therefore claims a consistent balance across criteria rather than strength on a single criterion (L355–356). Within that framing, HiOdex:

  • achieves the highest OA among the CAM-based methods and FovEx in all four settings (L351–352);
  • ranks among the top three in 22 of 24 model-oriented comparisons and never below fourth (L353–354);
  • remains within 0.0217 OA of the strongest perturbation-based result (L352–353), at approximately 1% of the computational cost (L92–93);
  • achieves the best or tied-best AUC-Judd in all four gaze settings (Table 2).

W2

"I would also expect some critique on the different metrics which are maybe not so suited : how to compare in a fait way CAM maps is not obvious."

We respectfully disagree that these points are missing: both are already addressed in the paper.

1/ Critique of the metrics: already in the paper. The model-oriented metrics capture complementary properties, and no method dominates them (L118–124).

EBPG, Sparsity, SIM and KL depend on the attribution values themselves, so they are sensitive to how concentrated a map is. The paper therefore reads Sparsity together with faithfulness and localization, because a concentrated map is not necessarily correct (App. B.5). The image-level gaze metrics cannot identify which object is explained, so they are complemented with object-level measures (L142–146).

2/ Fair comparison of CAM maps: already addressed by the evaluation protocol.

  • Same inputs. Every CAM-based method uses the same detector-specific target layers, so layer selection does not confound the comparison (L307–308). All methods explain the same detections under identical detector, input and target settings (L316–317).
  • Same scale. Every attribution map is normalized to [0,1][0, 1] before the model-oriented metrics are computed (App. B.5).
  • Scale-independent metrics. By their definitions, Insertion, Deletion and OA depend only on the order of attribution values (App. B.5), PG only on the location of the maximum (Eq. B1), and AUC-Judd only on how well attribution values separate fixated from non-fixated locations (App. B.6). These scores are therefore not affected by how a method scales its map.

Response to Reviewer pZUr (2/2)

Questions

"clarification requests : It is important to well define pretty much everything as your reader can come from different communities :"

Agreed. Key terms will be defined at first use (see Q1–Q2).

Q1

"line 73 not obvious for people not used to cam-like maps. For the maps we found attribution, the attribution maps without any definition of the attribution map. If this is obvious for someone working in XAI it is not for someone working in saliency for example."

Agreed. We will define "attribution map" at first use as a map that scores how strongly each image location supports the selected detection.

Q2

"73 a differentiable detector is defined later at line 201, again a term which should be briefly defined at first use. Again, obvious for XAI but not necessarily for others."

Agreed. We will move the definition to first use: a detector whose selected detection score is differentiable with respect to its feature layers, so gradients can be obtained by backpropagation (L201–205).

Q3

"120 I would add : "gradually" (or equivalent) revealed or removed. Also revealed and removed given what ? Given the importance in the map ?"

"Gradually": Agreed; we will add it at L120.

"Given what?": Yes, given the importance in the map. Deletion progressively removes pixels in descending attribution order, and Insertion restores them in the same order, 1,000 pixels per step (App. B.5).

Q4

"133 which and not whose if you refer to yolo ..."

Agreed. We will revise it to "YOLOv12-N, which retains real-time inference with an attention-centric design".

"work : Definitely what is needed is :"

Q5

"a discussion on the metrics stability (given different filtering techniques on the different maps for example)"

1/ Metric stability: already analyzed in the paper, in two places.

  • Human fixation maps: App. C.7 varies the fixation-density bandwidth σ\sigma and the clustering settings (Eps, MinPts). No configuration is uniformly optimal, so one fixed setting is applied to all methods (Fig. C15).
  • Attribution maps: App. C.5 varies the HiOdex smoothing width γ\gamma and reports its effect on PG, EBPG, OA and Sparsity (Fig. C7).

2/ Different filtering of the maps: we deliberately do not apply one. Each baseline keeps its published procedure (App. B.4), and applying an extra filter to all attribution maps would change the methods being compared.

Q6

"there are plenty of ways to extract potential fixations from saliency maps and several approaches might be taken into account"

Other ways to extract fixations: several are possible, but our choice is deliberate.

Why we keep ours: HiOdex uses evidence-driven fixation selection, which selects the strongest remaining evidence location at each step and then applies inhibition of return (L224–236). This choice has three advantages:

  • it is the human-inspired mechanism the method is built on (L73–76);
  • it uses only the detector's own evidence;
  • it needs no further detector evaluations (L285–286).

The results are also robust to this step. PG, OA and Sparsity remain approximately constant across the fixation-window width δ\delta, and all four metrics remain essentially unchanged across the number of fixations KK (App. C.4).

Q7

"the validation is convincing but there is no "wow""

We see this differently: the paper offers two distinctive contributions.

  • A new kind of explainer. HiOdex is the first post-hoc, human-inspired explainer that produces object-specific attributions for object detection without gaze supervision (L157–158).
  • A new kind of evaluation. To our knowledge, no detector explainer had been evaluated against human gaze at the dataset scale (L69–70).

With only one forward and one backward pass, HiOdex reaches faithfulness close to perturbation-based methods and the best or tied-best AUC-Judd against human gaze (see W1).

In summary, we will make the definition and wording fixes for Q1–Q4; the other points are already covered at the locations indicated. We hope these answers address the reviewer's comments and further strengthen their assessment.

Comments on the paper

4 comments

We thank the reviewers again for their constructive feedback. The revised version includes:

1/ Terminology (Reviewer pZUr): key terms are now defined at first use, and a new Appendix E explains specialized terms for readers from different communities.

2/ Human-gaze ablation (Reviewer GSJD): a new Appendix C.8 with Table C3 compares HiOdex with its Local Evidence Map alone, its Foveal CAM alone, and a single fixation on both gaze datasets. HiOdex outperforms all variants on every gaze metric.

AuthorsOpenReview ↗

REPORT OF REVIEWER qVm2 MISCONDUCT

To PC: WE WANT TO REPORT REVIEWER qVm2. The reviewer repeatedly claims gaps that our responses/paper have already addressed. They mention our replies but avoid the main points, shift the discussion to irrelevant details, and make claims that contradict one another. Our specific concerns remain unanswered, yet the reviewer continues to press the same objections.

NONE OF THE POINTS RAISED BY REVIEWER QVM2, INDIVIDUALLY OR TOGETHER, JUSTIFIES THE “CLEAR REJECTION” RECOMMENDATION. This pattern raises serious doubts about the reviewer’s technical judgment and good faith. It undermines the credibility of the review and obstructs a fair assessment of the paper.

PLEASE INVESTIGATE CAREFULLY AND BAN ALL INDIVIDUALS/INSTITUTES THAT ARE BEHIND THIS MISCONDUCT.

AuthorsOpenReview ↗

WITHDRAWAL REASON: MISCONDUCT AND UNACCEPTABLE REVIEW QUALITY OF REVIEWER qVm2

To PC, SAC, and AC:

WE WITHDRAW THIS SUBMISSION BECAUSE OF REVIEWER qVm2. This decision concerns Reviewer qVm2 only.

  1. Misconduct. The reviewer misrepresented prior work, ignored the evidence we provided, avoided our main points, and used unprofessional language about our method.
  2. Clear rejection despite agreement. The reviewer agreed with our key points during the discussion, yet kept the lowest rating ("1: Clear rejection") without identifying a single error in our method or results. A rating that does not change after such agreement is not based on the content of the paper. WE ASK THE PC TO INVESTIGATE WHAT IS BEHIND THIS RATING, INCLUDING ANY CONFLICT OF INTEREST.
  3. Subjective and dismissive assessment. The reviewer rated our paper "1: I learned little or nothing new from this paper", while the same review lists our human-gaze evaluation and our comparison across two datasets and two detectors as strengths. This rating is subjective, contradicts the reviewer's own assessment, and dismisses our contributions without any explanation.
  4. No constructive feedback. The review gave no constructive guidance on how the paper could be improved toward acceptance.

WE ARE DEEPLY UNSATISFIED WITH THE QUALITY OF REVIEWER qVm2's REVIEW. IT DOES NOT MEET THE STANDARD THAT ICLR EXPECTS FROM ITS REVIEWERS. FOR THIS REASON, WE WITHDRAW OUR SUBMISSION AND ASK THE PC TO INVESTIGATE REVIEWER qVm2 AND TAKE APPROPRIATE ACTION.

Reviewer_qVm2OpenReview ↗

My last comment, and I will not waste time to response anymore.

It is deeply disappointing that I have wasted so much time explaining my perspective and even offering sincere suggestions for improvement. The authors should have stood their ground—after all, they believe every point I raised is incorrect, and the assessment is subjective. A truly courageous action would have been to persist, subject the paper to formal review, and let the discussion be evaluated so that the more authoritative group could ultimately decide whether the paper should stand or fall. The authors, however, behaved like deserters.

--Reviewer qVm2

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.