acceptodds
Withdrawn from ICLR 2027

Evidential Curriculum Learning for Multi-Prompt Weakly Supervised RGB-T Salient Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-prompt weakly supervised RGB-Thermal salient object detection (WSSOD) reduces annotation costs with points, boxes, and scribbles, but suffers from unreliable predictions and unstable training caused by incomplete prompts, cross-modal inconsistency, and imbalanced sample difficulty. To address these issues, we propose an Evidential Curriculum Learning Network (ECL-Net), which builds a unified uncertainty-driven learning pipeline from region-level reliability modeling to sample-level curriculum optimization. Specifically, Evidential Certain-Uncertain Refinement (ECUR) mechanism models Dirichlet evidence to decouple confident and uncertain regions and refines ambiguous features with certainty-weighted prototypes and intra-class self-attention. Evidential Uncertainty Curriculum Module (EUCM) reuses this uncertainty estimate for easy-to-hard sample scheduling, enabling stable RGB-T saliency priors before harder cases. Extensive experiments under point-, box-, and scribble-supervised settings on VT821, VT1000, and VT5000 demonstrate that ECL-Net achieves state-of-the-art performance against existing RGB-T WSSOD methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
RGB-T salient object detection, multi-prompt weak supervision, Segment Anything Model 3, evidential deep learning, curriculum learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_c4j8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the problem in WSSOD that "exposing all samples equally from the start injects high-uncertainty gradients" and proposes the Evidential Certain–Uncertain Refinement mechanism. The mechanism decouples confident and uncertain regions via Dirichlet evidence, and refines ambiguous features with certainty-weighted prototypes and intra-class self-attention. The comparative experiments and visualization results demonstrate the effectiveness of the proposed method.

Critical strengths

The paper is well organized, clearly structured, and fluently written. The figures and tables are clear, making the paper easy to read and understand.

Critical weaknesses
  1. The paper claims to achieve state-of-the-art performance, but the results in Table 1 show that the proposed method underperforms SAMSOD on several metrics. This gap undermines the paper's claim to be state-of-the-art.
  2. The paper is concerned with the distinction between background (BG) and foreground (FG), yet it lacks visualization experiments that demonstrate the effectiveness of the proposed method in distinguishing BG from FG, which weakens the paper's novelty.
Questions for authors
  1. GN is not defined in Eq. (1). Does it refer to Group Normalization?
  2. As the core component of the proposed ECUR, EUCM appears to be presented through only a single equation (Eq. 12 in Section 3.3), while all the remaining equations are citations of existing work. This gives me some concerns regarding the novelty of the paper.
  3. In Table 1, the proposed method appears to fall clearly behind SAMSOD in the comparison. Does this weaken the novelty and effectiveness claimed for the paper?
  4. In Figure 4, the effect of ECUR does not look very pronounced. Have the authors considered comparing it with simple segmentation post-processing algorithms, such as Connected Component Filtering, which is commonly used in segmentation post-processing? Please explain the specific role of ECUR in conjunction with the visualization.
Reviewer_yX5rOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ECL-Net for weakly supervised RGB-T salient object detection. It covers point, box, and scribble supervision. The framework is built on SAM3. A prompt-guided decoder generates pseudo-labels, and a no-prompt decoder predicts saliency. The Evidential Certain-Uncertain Refinement mechanism (ECUR) estimates foreground and background evidence. It then selects reliable regions and refines features with certainty-weighted prototypes and class-masked attention. The Evidential Uncertainty Curriculum Module (EUCM) turns pixel uncertainty into a difficulty score for each sample. Training samples are then added from easy to hard. On VT821, VT1000, and VT5000, the method outperforms the listed weakly supervised baselines.

Critical strengths
  1. The motivation is clear. The paper addresses two problems: unreliable predictions within an image and uneven difficulty across images. The same uncertainty estimate is used for region refinement and for sample scheduling. This links the two modules in a natural way.
  2. The evaluation covers three RGB-T benchmarks and three annotation types. Table 2 gives module-level ablations. Figure 5 compares the evidential curriculum with an entropy-based one. These experiments go beyond simple comparisons with prior methods.
  3. The gains are large in some settings. On VT5000 with scribble supervision, ECL-Net raises the S-measure from 0.900 (SSFam) to 0.927 and lowers MAE from 0.028 to 0.021. EUCM also brings large improvements over the point- and box-supervised baselines.
Critical weaknesses
  1. Novelty and necessity are not well isolated. Prototype-guided correction, confidence-based region selection, class-restricted attention, and easy-to-hard training are all known techniques. SSFam already uses a similar prompt/no-prompt decoder design. The main contribution is likely the way these parts are combined through one evidential score. The experiments do not show that this score and this combination work better than simpler confidence-based refinement or standard sample scheduling. SAM3 is a strong pretrained prior, so controlled comparisons matter here.
  2. The meaning of ignorance and conflict is unclear. In the binary case, Equation (5) reduces to $
C=1−∣2p^fg−1∣=2min⁡(p^fg,1−p^fg)C=1-|2\hat p_{\mathrm{fg}}-1| =2\min(\hat p_{\mathrm{fg}},1-\hat p_{\mathrm{fg}})

$ So the conflict term depends only on the predictive mean. Also, when both class evidences go to zero, ignorance and conflict both approach one. This makes it hard to treat them as separate sources of uncertainty. The evidence head works on fused features and does not compare RGB and thermal evidence directly. The score alone therefore does not show disagreement between the two modalities. The combined score may still help, but its interpretation needs better support and a direct evaluation of uncertainty quality. 3. The gains from ECUR are small and uneven. Over B+EUCM, ECUR adds little. On VT5000 with scribble supervision, the S-measure rises from 0.925 to 0.927 and MAE stays at 0.021. On VT821 with scribble supervision, ECUR improves the S-measure but MAE gets worse, from 0.023 to 0.024. The paper reports no repeated runs or variance. It is hard to tell whether these small gains justify the extra complexity. The MAE gain under point supervision is clearer. The concern is about consistency and attribution, not about whether ECUR helps at all. 4. The large curriculum gains need clearer training controls. On VT5000, EUCM raises the point-supervised S-measure from 0.504 to 0.901 and the box-supervised S-measure from 0.749 to 0.902. These are very large jumps. However, the paper does not fully describe the warm-up size, validation split, plateau criterion, stage schedule, or total training budget. It is unclear how much of the gain comes from uncertainty ranking and how much comes from progressive data loading or different sample exposure. The entropy comparison in Figure 5 helps. A random-expansion baseline with the same number of optimization steps is also needed. In addition, Table 2 includes EUCM without ECUR, but EUCM relies on the evidential output. The paper should state which components and losses remain in this setting. 5. Several implementation and comparison details are missing. The supervision target and valid-pixel mask for the evidential loss are not given for each annotation type. Box supervision is the least clear case. Table 1 lists NESS-Net as an RGB-T box-supervised method, but the cited paper is about scribble-supervised RGB-D detection. Any adaptation should be described. Comparisons with RGB-only methods mix the effect of the extra modality with the effect of the method. Finally, the thresholds in Appendix A, �1��1�, match neither the stated search grid nor the constraint �2��2�.

Questions for authors
  1. What does the evidential score offer over plain confidence or entropy for seed selection in ECUR? Can the authors compare these options while keeping prototype refinement, attention, and losses fixed?
  2. The conflict term reduces to a function of the mean in the binary case. How should ignorance and conflict be read as separate quantities? What evidence links the estimated conflict to real disagreement between RGB and thermal inputs?
  3. Please give the full EUCM protocol: warm-up size, validation data and labels, plateau criterion, expansion rule, total optimization steps, and final data coverage. Do the gains hold against random expansion with the same training budget?
  4. What remains in B+EUCM when ECUR is removed? How do B, B+evi (Figure 5), and B+EUCM differ in evidence heads, losses, and feature refinement?
  5. Can the authors report results over repeated runs for the main comparisons? Can they also separate the effects of the evidential head and loss, prototype correction, and class-masked attention?
Reviewer_DvRfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies RGB and thermal salient object detection with point, box, and scribble supervision. It combines evidential feature refinement with a curriculum that introduces training samples according to predicted uncertainty. The method is evaluated on three test sets.

Critical strengths

The paper tackles a useful problem across three forms of weak supervision rather than a single prompt type. Table 2 also includes the SAM3 baseline and separate module ablations, which helps locate where the reported gains arise.

Critical weaknesses

The curriculum accounts for almost all of the improvement under box and point supervision. On VT821, the point baseline rises from 0.500 to 0.910 in S-measure with the curriculum alone. The paper does not establish whether this comes from evidential uncertainty, a simpler sample schedule, or instability in the baseline. It is also unclear how the curriculum obtains its uncertainty scores in the ablation without the evidential refinement module. The box comparison needs correction or explanation. Table 1 lists NESS-Net as an RGB and thermal method with box supervision, while the cited reference describes an RGB and depth method with scribble supervision. Appendix A adds to the reporting concerns: its selected threshold setting is not among the settings it says were tested. The paper also calls the model unified across prompt types there, whereas Table 1 describes separately trained models.

Questions for authors
  1. It is unclear whether the point, box and scribble results come from one unified model or three separately trained models. Table 1 describes models trained under each prompt type, while Appendix A calls the model unified. This distinction matters for the main claim.
  2. The curriculum produces an unusually large gain under point supervision. On VT821, S-measure rises from 0.500 for the baseline to 0.910 with EUCM alone. Please compare against random ordering and simpler curricula, and report results across several training runs.
  3. EUCM is said to reuse the uncertainty estimated by ECUR, yet Table 2 includes a baseline with EUCM and without ECUR. The paper should explain where its uncertainty scores come from in that setting.
  4. Table 1 lists NESS-Net as an RGB and thermal method with box supervision. The cited NESS-Net work is described in the references as an RGB and depth method with scribble supervision. If it was adapted and retrained, the procedure needs to be stated; otherwise this is not a valid box comparison.
  5. The point-supervised comparison is against PSOD, which uses RGB alone. A same-backbone RGB-only version of the proposed method would help show how much of the gain comes from the second modality rather than the proposed training scheme.
  6. The selected thresholds in Appendix A do not appear in the reported search grid. The adopted foreground and background thresholds differ by 0.55, although the text says that difference was fixed at 0.5. The adopted confidence threshold of 0.75 is also absent from the listed values.
  7. The SAM3 training procedure needs more detail. Please state which components are frozen, which are updated, and how prompts and pseudo-labels are generated for each supervision type. These choices could substantially affect the comparison.
  8. The paper uses uncertainty to guide both refinement and training order, but does not show whether high uncertainty actually identifies prediction errors. Such an analysis would be more persuasive than selected masks alone, particularly since the failure cases still show boundary errors.

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

Related papers

Discussion (0)

Sign in to comment.