We organize our supplementary material as follows: In Section A, we present more analytical results on VSPW dataset. In Section B, we provide more ablation studies on Cityscapes dataset. In Section D, we provide computational cost analysis and training details. In Section E, we provide comparison with bi-directional optical flow. A.1 Explanation of Video Consistency Following [3], we use Video Consistency (VC) to evaluate the category consistency among adjacent frames in the videos.

For out of distribution (OOD) inference, it is desired that the model can assign high epistemic uncertainty to the OOD regions compared to their ID counterparts. A.2 Policy Gradient based Reward Maximization for Segmentation Backbone This approach enables us to efficiently achieve the optimal solution for reward maximization. We present some examples of generated OOD examples in Figure 1(a). The results are presented in Figure 1(b)-(d). In Table 1, we present the results of our uncertainty estimation framework when applied to the Cityscapes dataset.

Whenever a merge operation is performed the corresponding edge is contracted and new edges can potentially be created (Lines 7-17). Afterwards, the clusters belonging to non-partitionable class (i.e stuff) are merged. Table 2 contains the hyperparameters used for fully differentiable training. We can see that optimizing PQ surrogate gives better performance and using separate losses decreases the performance especially on'thing' classes. Table 5 showing that all trials improve over the baseline by fully differentiable training.

Mohamed bin Zayed University of AI We organize our supplementary material as follows: In Section A, we present more analytical results on VSPW dataset. In Section B, we provide more ablation studies on Cityscapes dataset. In Section D, we provide computational cost analysis and training details. In Section E, we provide comparison with bi-directional optical flow. This observation demonstrates that our mask propagation framework implicitly captures the long-range temporal relationships among video frames.

Due to its ability to integrate information over large boxes, the new layer facilitates long-range propagation of information and leads to the efficient increase of the receptive fields of network units.

Depth imaging is a crucial area in Autonomous Driving Systems (ADS), as it plays a key role in detecting and measuring objects in the vehicle's surroundings. However, a significant challenge in this domain arises from missing information in Depth images, where certain points are not measurable due to gaps or inconsistencies in pixel data. Our research addresses two key tasks to overcome this challenge. First, we developed an algorithm using a multi-layered training approach to generate Depth images from a single RGB image. Second, we addressed the issue of missing information in Depth images by applying our algorithm to rectify these gaps, resulting in Depth images with complete and accurate data. We further tested our algorithm on the Cityscapes dataset and successfully resolved the missing information in its Depth images, demonstrating the effectiveness of our approach in real-world urban environments.