CoLA: Conditional Dropout and Language-driven Robust Dual-modal Salient Object Detection

TL;DR

CoLA tackles robust dual-modal salient object detection under noisy inputs and modality missing by integrating a language-driven quality assessment (LQA) with a prompt-tuned vision-language model and a conditional dropout (CD) training scheme. Stage I trains LQA to reweight modality contributions using fixed prompts and a learnable prompt in CLIP, producing quality scores that modulate feature fusion via cosine similarity-based fusion. Stage II employs CD by duplicating encoders, freezing the original, and training a zero-convolution-augmented copy under missing-modality conditions, ensuring performance under both complete and incomplete inputs. The method achieves state-of-the-art results on RGB-T and RGB-D benchmarks across modality-complete and modality-missing scenarios, with ablations validating the contributions of LQA and CD. The approach is simple, scalable, and readily extensible to other multi-modal perception tasks, with code to be released upon acceptance.

Abstract

The depth/thermal information is beneficial for detecting salient object with conventional RGB images. However, in dual-modal salient object detection (SOD) model, the robustness against noisy inputs and modality missing is crucial but rarely studied. To tackle this problem, we introduce \textbf{Co}nditional Dropout and \textbf{LA}nguage-driven(\textbf{CoLA}) framework comprising two core components. 1) Language-driven Quality Assessment (LQA): Leveraging a pretrained vision-language model with a prompt learner, the LQA recalibrates image contributions without requiring additional quality annotations. This approach effectively mitigates the impact of noisy inputs. 2) Conditional Dropout (CD): A learning method to strengthen the model's adaptability in scenarios with missing modalities, while preserving its performance under complete modalities. The CD serves as a plug-in training scheme that treats modality-missing as conditions, strengthening the overall robustness of various dual-modal SOD models. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art dual-modal SOD models, under both modality-complete and modality-missing conditions. We will release source code upon acceptance.

Figures (9)

• Figure 1: (a): The two examples demonstrate the cases when the RGB or the Depth inputs are noisy, QA means quality assessment. (b): examples of two modality-missing conditions. The proposed CoLA produces robust results.
• Figure 1: Different cases of our LQA, $\beta_{rgb} = \frac{\alpha^{rgb}}{\alpha^{rgb}+\alpha^{t}}$,$\beta_{t} = \frac{\alpha^{t}}{\alpha^{rgb}+\alpha^{t}}$.
• Figure 2: The architecture of CoLA represents a two-stage neural network with Stage I training a language quality assessment (LQA) to calibrate feature fusion, and Stage II training with Conditional Dropout enhances the capabilities of both missing and complete modalities.
• Figure 2: Comparison of Models Trained with Conditional Dropout, Modality Dropout, and Original Model in VT1000 and VT821 datasets.
• Figure 3: Architectural comparison of (a) No-Reference Method, (b) Pre-trained Assessment and (c) Our LQA.