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 down-sampling 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.

As neural networks grow in size and complexity, inference speeds decline. To combat this, one of the most effective compression techniques -- channel pruning -- removes channels from weights. However, for multi-branch segments of a model, channel removal can introduce inference-time memory copies. In turn, these copies increase inference latency -- so much so that the pruned model can be slower than the unpruned model. As a workaround, pruners conventionally constrain certain channels to be pruned together. This fully eliminates memory copies but, as we show, significantly impairs accuracy. We now have a dilemma: Remove constraints but increase latency, or add constraints and impair accuracy. In response, our insight is to reorder channels at export time, (1) reducing latency by reducing memory copies and (2) improving accuracy by removing constraints. Using this insight, we design a generic algorithm UPSCALE to prune models with any pruning pattern. By removing constraints from existing pruners, we improve ImageNet accuracy for post-training pruned models by 2.1 points on average -- benefiting DenseNet (+16.9), EfficientNetV2 (+7.9), and ResNet (+6.2). Furthermore, by reordering channels, UPSCALE improves inference speeds by up to 2x over a baseline export.

Polarimetric SAR (PolSAR) is uniquely able to capture structural and compositional properties of targets leading to improved performance in various classification applications over traditional single-polarization SAR. To aid in the interpretation of the return scatter, several decomposition techniques have been developed that attempt to classify the scene by presenting the return power as a combination of pre-determined canonical targets. For most decomposition techniques, these relationships are derived from the physics of radar scatter, and thus the results of the segmentation are explainable in terms of observable physical phenomena. Recently, Deep Neural Networks (DNNs) have emerged as a leading strategy for classifying PolSAR data. While effective, these techniques are dependent on discovering relationships between the data and provided supervised labels during an exploratory phase of the algorithm called “training”. The inter-dependencies are embedded into the weights of the network and are subsequently used to perform classification of the unlabeled data samples. Since the entire process is data-driven, it can be difficult to explain the outcomes of the algorithm physically. In this paper, we begin to explore the relationship between radar physics and the latent space embedded in the networks during supervised training. The goal is to explore current and future possibilities of explaining the results of deep neural networks and relating their outputs to radar physics. This can help improve confidence in the outputs of DNNs and potentially illuminate strategies to further improve their performance by embedding physical constraints in their training and classification.

RISAT-1 data is acquired over Mumbai in hybrid and linear dual polarizations. The mean and standard deviation of backscattering coefficients ( σ0) are computed and have been analyzed for various land features. Classification accuracy between RISAT-1 hybrid and dual polarimetric data has been compared. The effect of different multilook on the classification accuracy is also studied. Wishart supervised and Support Vector Machine (SVM) classifiers are used for this study. It has been observed that the classification accuracy can be improved by using m-δ or m-χ decomposition along with Circular Polarization Ratio (CPR) and SPAN of hybrid polarimetric data.