- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper starts from an interesting tension: context can be harmful when it acts as a shortcut, but useful when direct object evidence is missing. The paper proposes Evidence-Gated Backdoor Adjustment (EG-BA), a module for transformer-based open-set object detection. The motivating hypothesis is that an object query mixes direct visual evidence with scene context. The authors argue that the same context produces two opposing effects: it can induce co-occurrence shortcuts when the object is visible, while helping recognition when the object is occluded.
Critical strengths
The motivating observation about the value and problem of context is worthwhile. Framing context as conditionally useful rather than uniformly harmful is nuanced and makes the paper have an interesting prelude.
The authors report important negative findings, they don't hide them. Some things didn't work and they report it. E.g. gate collapse under the information-bottleneck term, lack of transfer to natural occlusion, and failure on the co-occurrence-shift test. This transparency is valuable.
The ablations include relevant variations on the topic. E.g. evidence supervision, threshold sensitivity, gate distributions, and synthetic occlusion severity. These analyses expose failure modes.
Critical weaknesses
No solid evidence is provided that the learned dictionary identifies confounders rather than generic feature prototypes. That would have been very valuable. But there is no intervention, prototype visualization, or controlled experiment to demonstrate its effects.
The information-bottleneck is presented as a contribution, but it collapses the gate and is set to zero in the final model. The claim of the authors is that an adaptive balance is needed between debiasing and context recovery, but the context-only performs best.
The method fails on co-occurrence bias and natural occlusion, which it target explicitly.
The paper lists Don’t Judge an Object by Its Context, but it does not discuss it seriously, neither do they compare against it. Causal attention and context-debiasing methods are discussed but not taken as baselines.
Occlusion-aware literature seem to be a few and mainly older papers. The paper would be stronger when modern transformer-based approaches are discussed as well.
Questions for authors
What is the relation and difference to "Explicit Modeling of Causal Factors and Confounders for Image Classification", AAAI 2026 by Wei Wu, et al.?
Can you provide a matched non-causal gating baseline with the same architecture and parameter count, so the contribution of the gating can be emphasized?
Can you compare against context-debiasing and occlusion-aware methods under the same Deformable DETR backbone and training protocol?
2 comments
Response to Rebuttal
Dear authors, I appreciate the honesty and openness in your responses. It helps to better understand this paper's choices, limitations, and implications. It does not cause me to change my score. My reasoning is as follows: most of the problems were acknowledged by the authors (thank you), some of them were addressed in textual changes (e.g. scope and precision of claims, related works, more precisely defined contributions), but my main concerns were not resolved during the rebuttal. It's fair of the authors to say that additional experiments regarding very related works as baselines, were not feasible in the rebuttal period. The authors' confirmation about limitations of the method, and that baseline experiments would have been helpful, confirms my earlier assessment that the paper does not reach ICLR quality at this stage.
Response to Reviewer td5e: dictionary analysis, baselines, limitations
Thank you for the detailed review, and for recognising the negative findings. All new numbers come from one training run per configuration (single seed); we flag where this limits the conclusion.
W1 (no evidence that the dictionary identifies confounders). Fair. We ran a prototype analysis on 3,894 confident predictions (K=20): attention is diffuse (entropy 2.75 vs max 3.00; mean max weight 0.15 vs 0.05 uniform), all 20 prototypes are used, and NMI between the top prototype and predicted class is 0.16. We will not claim the prototypes are semantic confounders and will describe the dictionary as a learned context basis. We have no intervention experiment.
W2 (IB as a contribution; context-only performs best). Agreed on both. At lambda_IB=0.1 the gate collapses to a constant (mean 0.0179, std 0.0000); the headline uses lambda_IB=0 plus a small anti-collapse term. We will reframe IB as a documented negative finding. Fixed e=0 has higher mAP (52.82 vs 51.20) and U-Recall (0.0689 vs 0.0573); the adaptive gate is better on WI and A-OSE. We claim only this trade-off.
W3 (fails on co-occurrence and natural occlusion). Agreed: Table 7 shows parity and Table 8 shows no meaningful gap reduction. We will say in the abstract and conclusion that we do not demonstrate bias mitigation or natural-occlusion gains. 8 of 10 co-occurrence pairs involve person. Re-analysis. Table 8's EG-BA entry (-0.1344) came from the collapsed lambda_IB=0.1 checkpoint, not the lambda_IB=0 headline model; this was our error. For the headline model the mean gap is -0.1121 vs -0.1186 for vanilla (matched MLP variant: -0.1185). Excluding person, pairs re-mined from the remaining classes (10 pairs) give -0.0736 (EG-BA), -0.0867 (vanilla), -0.0703 (MLP variant). So excluding person shrinks the gap by 27-41% depending on the model, consistent with an occlusion/crowding effect, but a negative gap remains in all models, EG-BA is only 0.013 narrower than vanilla, and the MLP variant is equally narrow. Negative pairs: 8/10 in all three models (the submission's 9/10 was a miscount) and 7/10, 6/10, 6/10 without person. We therefore do not claim bias mitigation.
W4 (Don't Judge an Object by Its Context; causal baselines). We have not compared against CONTA/CATT-style methods and could not reproduce them under our protocol in the rebuttal period. Singh et al. (2020) decorrelates a category's features from its co-occurring context via a feature split for multi-label image classification; it works at the image level with no per-query evidence signal and no occlusion recovery. We will discuss it properly in Related Work; applying a feature split to detection queries would be a natural baseline that we have not run.
W5 (occlusion literature is old). We will add recent transformer-based occlusion work.
Q1 (relation to Wu et al., AAAI 2026). We will read and discuss it in the revision; we do not want to claim a difference we have not verified.
Q2 (matched non-causal gating baseline). Done: same gate, loss, recipe and data; the dictionary attention is replaced by an MLP of the query with matched context parameters (268,042 vs 268,288). Result (single seed): the MLP variant is worse on known mAP (46.90 vs 51.20), WI (0.0193 vs 0.0097) and A-OSE (1417 vs 1224), and better on U-Recall (0.0767 vs 0.0573). Because the other re-run configurations (K=10, 50) also fall 5-7 mAP points below our K=20 run, seed variance may be as large as these gaps, so we do not claim the dictionary beats a generic MLP; we will report this as a mixed result. The full module adds 0.565M parameters (28.70M vs 28.13M), which we had wrongly reported as zero.
Q3 (comparison under the same backbone). Not available within the rebuttal period, as above. Table 9 compares against published numbers from other protocols; we will state this and not claim superiority.
Also found: Table 2 (K ablation) used lambda_IB=0.1; we re-ran K=10/50 at lambda_IB=0. K ablation re-run at lambda_IB=0 (mAP / WI / A-OSE / U-Recall avg): K=10: 45.94 / 0.0174 / 1349 / 0.0865; K=20: 51.20 / 0.0097 / 1224 / 0.0573; K=50: 44.10 / 0.0075 / 1319 / 0.0887. WI falls as K grows, U-Recall rises with K, and K=20 is best on mAP and A-OSE (single seed). We will replace the Table 2 discussion with these runs, keeping Table 2 as an explicitly labelled lambda_IB=0.1 result.