Test-Time Intensity Consistency Adaptation for Shadow Detection

TL;DR

TICA is introduced, a novel framework that leverages light-intensity information during test-time adaptation to enhance shadow detection accuracy and outperforms existing state-of-the-art methods, achieving superior results in balanced error rate (BER).

Abstract

Shadow detection is crucial for accurate scene understanding in computer vision, yet it is challenged by the diverse appearances of shadows caused by variations in illumination, object geometry, and scene context. Deep learning models often struggle to generalize to real-world images due to the limited size and diversity of training datasets. To address this, we introduce TICA, a novel framework that leverages light-intensity information during test-time adaptation to enhance shadow detection accuracy. TICA exploits the inherent inconsistencies in light intensity across shadow regions to guide the model toward a more consistent prediction. A basic encoder-decoder model is initially trained on a labeled dataset for shadow detection. Then, during the testing phase, the network is adjusted for each test sample by enforcing consistent intensity predictions between two augmented input image versions. This consistency training specifically targets both foreground and background intersection regions to identify shadow regions within images accurately for robust adaptation. Extensive evaluations on the ISTD and SBU shadow detection datasets reveal that TICA significantly demonstrates that TICA outperforms existing state-of-the-art methods, achieving superior results in balanced error rate (BER).

Figures (5)

• Figure 1: Light intensity inconsistency in shadow detection. Row two displays the results from the pre-trained model. Without the TTA strategy, the model tends to identify dark areas as shadow regions. Following the application of our TICA strategy, the prediction masks show improved accuracy in detecting shadows.
• Figure 2: Overview of the proposed TICA. By leveraging light consistency training, the TICA framework enhances the model's capabilities in shadow detection. Initially, the model is trained with a publicly accessible shadow detection dataset. We then apply random data augmentation techniques—horizontal flipping, resizing, and cropping—to the test set. This facilitates model refinement by enforcing consistent intensity predictions between the two augmented images. The consistency loss is backpropagated to update the encoder.
• Figure 3: The impact of the proposed TICA fluctuates with the number of epochs. Our TICA strategy is evaluated on the ISTD and SBU datasets across three backbone architectures: ResNet-50, Swin-Tiny, and HRNet-18.
• Figure 4: Qualitative comparison of the fine-tuning results and ground truth for TICA over five epochs. We observe that the original prediction masks from the pre-trained model are coarse, with results improving as the TTA epochs increase (within 5 epochs).
• Figure 5: Visual comparison of other SOTA Shadow Detectors and our method (TICA) against ground truth's shadow mask. It is evident that our method can detect shadow regions more accurately. Our results in the fifth column exhibit greater consistency with the ground truth in the sixth column than the other methods shown in the second to fourth columns on the ISTD dataset.