Philipe Dias

Philipe Dias, Aristeidis Tsaris, Jordan Bowman et al.

While the pretraining of Foundation Models (FMs) for remote sensing (RS) imagery is on the rise, models remain restricted to a few hundred million parameters. Scaling models to billions of parameters has been shown to yield unprecedented benefits including emergent abilities, but requires data scaling and computing resources typically not available outside industry R&D labs. In this work, we pair high-performance computing resources including Frontier supercomputer, America's first exascale system, and high-resolution optical RS data to pretrain billion-scale FMs. Our study assesses performance of different pretrained variants of vision Transformers across image classification, semantic segmentation and object detection benchmarks, which highlight the importance of data scaling for effective model scaling. Moreover, we discuss construction of a novel TIU pretraining dataset, model initialization, with data and pretrained models intended for public release. By discussing technical challenges and details often lacking in the related literature, this work is intended to offer best practices to the geospatial community toward efficient training and benchmarking of larger FMs.

Aristeidis Tsaris, Philipe Ambrozio Dias, Abhishek Potnis et al.

As AI workloads increase in scope, generalization capability becomes challenging for small task-specific models and their demand for large amounts of labeled training samples increases. On the contrary, Foundation Models (FMs) are trained with internet-scale unlabeled data via self-supervised learning and have been shown to adapt to various tasks with minimal fine-tuning. Although large FMs have demonstrated significant impact in natural language processing and computer vision, efforts toward FMs for geospatial applications have been restricted to smaller size models, as pretraining larger models requires very large computing resources equipped with state-of-the-art hardware accelerators. Current satellite constellations collect 100+TBs of data a day, resulting in images that are billions of pixels and multimodal in nature. Such geospatial data poses unique challenges opening up new opportunities to develop FMs. We investigate billion scale FMs and HPC training profiles for geospatial applications by pretraining on publicly available data. We studied from end-to-end the performance and impact in the solution by scaling the model size. Our larger 3B parameter size model achieves up to 30% improvement in top1 scene classification accuracy when comparing a 100M parameter model. Moreover, we detail performance experiments on the Frontier supercomputer, America's first exascale system, where we study different model and data parallel approaches using PyTorch's Fully Sharded Data Parallel library. Specifically, we study variants of the Vision Transformer architecture (ViT), conducting performance analysis for ViT models with size up to 15B parameters. By discussing throughput and performance bottlenecks under different parallelism configurations, we offer insights on how to leverage such leadership-class HPC resources when developing large models for geospatial imagery applications.

Philipe A. Dias, Henry Medeiros

Semantic segmentation with fine-grained pixel-level accuracy is a fundamental component of a variety of computer vision applications. However, despite the large improvements provided by recent advances in the architectures of convolutional neural networks, segmentations provided by modern state-of-the-art methods still show limited boundary adherence. We introduce a fully unsupervised post-processing algorithm that exploits Monte Carlo sampling and pixel similarities to propagate high-confidence pixel labels into regions of low-confidence classification. Our algorithm, which we call probabilistic Region Growing Refinement (pRGR), is based on a rigorous mathematical foundation in which clusters are modelled as multivariate normally distributed sets of pixels. Exploiting concepts of Bayesian estimation and variance reduction techniques, pRGR performs multiple refinement iterations at varied receptive fields sizes, while updating cluster statistics to adapt to local image features. Experiments using multiple modern semantic segmentation networks and benchmark datasets demonstrate the effectiveness of our approach for the refinement of segmentation predictions at different levels of coarseness, as well as the suitability of the variance estimates obtained in the Monte Carlo iterations as uncertainty measures that are highly correlated with segmentation accuracy.

Philipe A. Dias, Damiano Malafronte, Henry Medeiros et al.

Effective assisted living environments must be able to perform inferences on how their occupants interact with one another as well as with surrounding objects. To accomplish this goal using a vision-based automated approach, multiple tasks such as pose estimation, object segmentation and gaze estimation must be addressed. Gaze direction in particular provides some of the strongest indications of how a person interacts with the environment. In this paper, we propose a simple neural network regressor that estimates the gaze direction of individuals in a multi-camera assisted living scenario, relying only on the relative positions of facial keypoints collected from a single pose estimation model. To handle cases of keypoint occlusion, our model exploits a novel confidence gated unit in its input layer. In addition to the gaze direction, our model also outputs an estimation of its own prediction uncertainty. Experimental results on a public benchmark demonstrate that our approach performs on pair with a complex, dataset-specific baseline, while its uncertainty predictions are highly correlated to the actual angular error of corresponding estimations. Finally, experiments on images from a real assisted living environment demonstrate the higher suitability of our model for its final application.

Philipe A. Dias, Amy Tabb, Henry Medeiros

In fruit production, critical crop management decisions are guided by bloom intensity, i.e., the number of flowers present in an orchard. Despite its importance, bloom intensity is still typically estimated by means of human visual inspection. Existing automated computer vision systems for flower identification are based on hand-engineered techniques that work only under specific conditions and with limited performance. This work proposes an automated technique for flower identification that is robust to uncontrolled environments and applicable to different flower species. Our method relies on an end-to-end residual convolutional neural network (CNN) that represents the state-of-the-art in semantic segmentation. To enhance its sensitivity to flowers, we fine-tune this network using a single dataset of apple flower images. Since CNNs tend to produce coarse segmentations, we employ a refinement method to better distinguish between individual flower instances. Without any pre-processing or dataset-specific training, experimental results on images of apple, peach and pear flowers, acquired under different conditions demonstrate the robustness and broad applicability of our method.

Philipe A. Dias, Amy Tabb, Henry Medeiros

To optimize fruit production, a portion of the flowers and fruitlets of apple trees must be removed early in the growing season. The proportion to be removed is determined by the bloom intensity, i.e., the number of flowers present in the orchard. Several automated computer vision systems have been proposed to estimate bloom intensity, but their overall performance is still far from satisfactory even in relatively controlled environments. With the goal of devising a technique for flower identification which is robust to clutter and to changes in illumination, this paper presents a method in which a pre-trained convolutional neural network is fine-tuned to become specially sensitive to flowers. Experimental results on a challenging dataset demonstrate that our method significantly outperforms three approaches that represent the state of the art in flower detection, with recall and precision rates higher than $90\%$. Moreover, a performance assessment on three additional datasets previously unseen by the network, which consist of different flower species and were acquired under different conditions, reveals that the proposed method highly surpasses baseline approaches in terms of generalization capability.

Philipe A. Dias, Henry Medeiros

Despite recent improvements using fully convolutional networks, in general, the segmentation produced by most state-of-the-art semantic segmentation methods does not show satisfactory adherence to the object boundaries. We propose a method to refine the segmentation results generated by such deep learning models. Our method takes as input the confidence scores generated by a pixel-dense segmentation network and re-labels pixels with low confidence levels. The re-labeling approach employs a region growing mechanism that aggregates these pixels to neighboring areas with high confidence scores and similar appearance. In order to correct the labels of pixels that were incorrectly classified with high confidence level by the semantic segmentation algorithm, we generate multiple region growing steps through a Monte Carlo sampling of the seeds of the regions. Our method improves the accuracy of a state-of-the-art fully convolutional semantic segmentation approach on the publicly available COCO and PASCAL datasets, and it shows significantly better results on selected sequences of the finely-annotated DAVIS dataset.