AutoFocusFormer: Image Segmentation off the Grid

Real world images often have highly imbalanced content density. Some areas are very uniform, e.g., large patches of blue sky, while other areas are scattered with many small objects. Yet, the commonly used successive grid downsampling strategy in convolutional deep networks treats all areas equally. Hence, small objects are represented in very few spatial locations, leading to worse results in tasks such as segmentation. Intuitively, retaining more pixels representing small objects during downsampling helps to preserve important information. To achieve this, we propose AutoFocusFormer (AFF), a local-attention transformer image recognition backbone, which performs adaptive downsampling by learning to retain the most important pixels for the task. Since adaptive downsampling generates a set of pixels irregularly distributed on the image plane, we abandon the classic grid structure. Instead, we develop a novel point-based local attention block, facilitated by a balanced clustering module and a learnable neighborhood merging module, which yields representations for our point-based versions of state-of-the-art segmentation heads. Experiments show that our AutoFocusFormer (AFF) improves significantly over baseline models of similar sizes.

Improved Point Transformation Methods For Self-Supervised Depth Prediction

Given stereo or egomotion image pairs, a popular and successful method for unsupervised learning of monocular depth estimation is to measure the quality of image reconstructions resulting from the learned depth predictions. Continued research has improved the overall approach in recent years, yet the common framework still suffers from several important limitations, particularly when dealing with points occluded after transformation to a novel viewpoint. While prior work has addressed this problem heuristically, this paper introduces a z-buffering algorithm that correctly and efficiently handles occluded points. Because our algorithm is implemented with operators typical of machine learning libraries, it can be incorporated into any existing unsupervised depth learning framework with automatic support for differentiation. Additionally, because points having negative depth after transformation often signify erroneously shallow depth predictions, we introduce a loss function to penalize this undesirable behavior explicitly. Experimental results on the KITTI data set show that the z-buffer and negative depth loss both improve the performance of a state of the art depth-prediction network.

Visualizing Point Cloud Classifiers by Curvature Smoothing

Recently, several networks that operate directly on point clouds have been proposed. There is significant utility in understanding them better, so that humans can understand more about the mechanisms how those networks classify point clouds, potentially helping diagnosing them and designing better architectures and data augmentation pipelines. In this paper, we propose a novel approach to visualize important features used in classification decisions of point cloud networks. Following ideas in visualizing 2-D convolutional networks, our approach is based on gradually smoothing parts of the point cloud. However, different from the 2-D case, we smooth the curvature of the point cloud to remove sharp shape features. The resulting point cloud is then evaluated on the original point cloud network to see whether the performance has dropped or remained the same, from which parts that are important to the point cloud classification are identified. A technical contribution of the paper is an approximated curvature smoothing algorithm, which can smoothly transition from the original point cloud to one of constant curvature, such as a uniform sphere. With this smoothing algorithm, we propose PCI-GOS, a 3-D extension of the Integrated-Gradients Optimized Saliency (I-GOS) algorithm, as a perturbation-based visualization technique realized on 3-D shapes. Experiment results revealed insights into these classifiers.