PointGoal navigation has seen significant recent interest and progress, spurred on by the Habitat platform and associated challenge. In this paper, we study PointGoal navigation under both a sample budget (75 million frames) and a compute budget (1 G
PU for 1 day). We conduct an extensive set of experiments, cumulatively totaling over 50,000 GPU-hours, that let us identify and discuss a number of ostensibly minor but significant design choices -- the advantage estimation procedure (a key component in training), visual encoder architecture, and a seemingly minor hyper-parameter change. Overall, these design choices to lead considerable and consistent improvements over the baselines present in Savva et al. Under a sample budget, performance for RGB-D agents improves 8 SPL on Gibson (14% relative improvement) and 20 SPL on Matterport3D (38% relative improvement). Under a compute budget, performance for RGB-D agents improves by 19 SPL on Gibson (32% relative improvement) and 35 SPL on Matterport3D (220% relative improvement). We hope our findings and recommendations will make serve to make the communitys experiments more efficient.
Graphic design is essential for visual communication with layouts being fundamental to composing attractive designs. Layout generation differs from pixel-level image synthesis and is unique in terms of the requirement of mutual relations among the de
sired components. We propose a method for design layout generation that can satisfy user-specified constraints. The proposed neural design network (NDN) consists of three modules. The first module predicts a graph with complete relations from a graph with user-specified relations. The second module generates a layout from the predicted graph. Finally, the third module fine-tunes the predicted layout. Quantitative and qualitative experiments demonstrate that the generated layouts are visually similar to real design layouts. We also construct real designs based on predicted layouts for a better understanding of the visual quality. Finally, we demonstrate a practical application on layout recommendation.
Robot perception systems need to perform reliable image segmentation in real-time on noisy, raw perception data. State-of-the-art segmentation approaches use large CNN models and carefully constructed datasets; however, these models focus on accuracy
at the cost of real-time inference. Furthermore, the standard semantic segmentation datasets are not large enough for training CNNs without augmentation and are not representative of noisy, uncurated robot perception data. We propose improving the performance of real-time segmentation frameworks on robot perception data by transferring features learned from synthetic segmentation data. We show that pretraining real-time segmentation architectures with synthetic segmentation data instead of ImageNet improves fine-tuning performance by reducing the bias learned in pretraining and closing the textit{transfer gap} as a result. Our experiments show that our real-time robot perception models pretrained on synthetic data outperform those pretrained on ImageNet for every scale of fine-tuning data examined. Moreover, the degree to which synthetic pretraining outperforms ImageNet pretraining increases as the availability of robot data decreases, making our approach attractive for robotics domains where dataset collection is hard and/or expensive.
We propose a self-supervised learning method to jointly reason about spatial and temporal context for video recognition. Recent self-supervised approaches have used spatial context [9, 34] as well as temporal coherency [32] but a combination of the t
wo requires extensive preprocessing such as tracking objects through millions of video frames [59] or computing optical flow to determine frame regions with high motion [30]. We propose to combine spatial and temporal context in one self-supervised framework without any heavy preprocessing. We divide multiple video frames into grids of patches and train a network to solve jigsaw puzzles on these patches from multiple frames. So the network is trained to correctly identify the position of a patch within a video frame as well as the position of a patch over time. We also propose a novel permutation strategy that outperforms random permutations while significantly reducing computational and memory constraints. We use our trained network for transfer learning tasks such as video activity recognition and demonstrate the strength of our approach on two benchmark video action recognition datasets without using a single frame from these datasets for unsupervised pretraining of our proposed video jigsaw network.
Adverse surgical outcomes are costly to patients and hospitals. Approaches to benchmark surgical care are often limited to gross measures across the entire procedure despite the performance of particular tasks being largely responsible for undesirabl
e outcomes. In order to produce metrics from tasks as opposed to the whole procedure, methods to recognize automatically individual surgical tasks are needed. In this paper, we propose several approaches to recognize surgical activities in robot-assisted minimally invasive surgery using deep learning. We collected a clinical dataset of 100 robot-assisted radical prostatectomies (RARP) with 12 tasks each and propose `RP-Net, a modified version of InceptionV3 model, for image based surgical activity recognition. We achieve an average precision of 80.9% and average recall of 76.7% across all tasks using RP-Net which out-performs all other RNN and CNN based models explored in this paper. Our results suggest that automatic surgical activity recognition during RARP is feasible and can be the foundation for advanced analytics.
Purpose: Manual feedback in basic RMIS training can consume a significant amount of time from expert surgeons schedule and is prone to subjectivity. While VR-based training tasks can generate automated score reports, there is no mechanism of generati
ng automated feedback for surgeons performing basic surgical tasks in RMIS training. In this paper, we explore the usage of different holistic features for automated skill assessment using only robot kinematic data and propose a weighted feature fusion technique for improving score prediction performance. Methods: We perform our experiments on the publicly available JIGSAWS dataset and evaluate four different types of holistic features from robot kinematic data - Sequential Motion Texture (SMT), Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT) and Approximate Entropy (ApEn). The features are then used for skill classification and exact skill score prediction. Along with using these features individually, we also evaluate the performance using our proposed weighted combination technique. Results: Our results demonstrate that these holistic features outperform all previous HMM based state-of-the-art methods for skill classification on the JIGSAWS dataset. Also, our proposed feature fusion strategy significantly improves performance for skill score predictions achieving up to 0.61 average spearman correlation coefficient. Conclusions: Holistic features capturing global information from robot kinematic data can successfully be used for evaluating surgeon skill in basic surgical tasks on the da Vinci robot. Using the framework presented can potentially allow for real time score feedback in RMIS training.
