Map Pruning Research Articles

Landmark-aware autonomous odometry correction and map pruning for planetary rovers

Planetary rover autonomous localization is paramount for a planetary surface exploration mission. However, existing methods demonstrate limited localization accuracy, mostly due to the unstructured texture characterization of planetary surface. In response, this study presents a novel Neural Radiance Field (NeRF) driven visual odometry correction method that allows for high-precision 6-DoF rover pose estimation and local map pruning. First, an innovative image saliency evaluation approach, combining binarization and feature detection, is introduced to meticulously select landmarks that are conducive to rover re-localization. Subsequently, we conduct 3D reconstruction and rendering of the chosen landmarks based on a-priori knowledge of planetary surface images and their Neural Radiance Field (NeRF) models. High-precision odometry correction is achieved through the optimization of photometric loss between NeRF rending images and real images. Simultaneously, the odometry correction mechanism is employed in an autonomous manner to refine the NeRF model of the corresponding landmark, leading to an improved local map and gradually enhanced rover localization accuracy. Numerical simulation and experiment trials are carried out to evaluate the performance of the proposed method, results of which demonstrate state-of-the-art rover re-localization accuracy and local map pruning.

AntiDoteX: Attention-Based Dynamic Optimization for Neural Network Runtime Efficiency

Deep neural networks (DNNs) achieved great cognitive performance at the expense of a considerable computation workload. To relieve the computational burden, many optimization works are developed to reduce the model redundancy by identifying and removing insignificant model components, such as weight sparsity and filter pruning methods. However, these works only evaluate model components’ static significance with parameter information, ignoring their dynamic interaction with external inputs. Specifically, due to the difference in per-input features, the model components’ significance can dynamically change and, thus, the static methods can only achieve suboptimal performance. Focusing on this aspect, we propose a dynamic DNN optimization framework in this work. Based on the neural network attention mechanism, we propose a comprehensive dynamic optimization framework, including 1) testing-phase dynamic feature map pruning; 2) training-phase optimization by training with targeted dropout; and 3) deployment-phase one-for-all (OFA) model adaptability enhancement. By providing a holistic dynamic testing, training, and deployment co-optimization framework, our work has the following benefits: first, it can accurately identify and aggressively remove per-input feature redundancy by considering the model-input interaction and involving the channel/column-wise pruning flexibility; meanwhile, the training-testing co-optimization favors the dynamic pruning and helps maintain the model accuracy even with a very high feature pruning ratio. Finally, the deployment enhancement provides one unified OFA model to support full-spectrum feature sparsity ratios. The unified model can be dynamically reconfigured to meet different resource budgets without any retraining cost, and thus provide significant deployment flexibility. Extensive experiments show that our method could bring 37.4%–54.5% floating-point operations reduction with negligible accuracy drop on various test benchmarks. Meanwhile, the OFA deployment optimization enables us to use one model to support at most ten different resource constraints without any retraining cost.

Dynamic Regularization on Activation Sparsity for Neural Network Efficiency Improvement

When deploying deep neural networks in embedded systems, it is crucial to decrease the model size and computational complexity for improving the execution speed and efficiency. In addition to conventional compression techniques, e.g., weight pruning and quantization, removing unimportant activations can also dramatically reduce the amount of data communication and the computation cost. Unlike weight parameters, the pattern of activations is directly related to input data and thereby changes dynamically. To regulate the dynamic activation sparsity (DAS), in this work, we propose a generic low-cost approach based on winners-take-all (WTA) dropout technique. The network enhanced by the proposed WTA dropout, namely DASNet , features structured activation sparsity with an improved sparsity level. Compared to the static feature map pruning methods, DASNets provide better computation cost reduction. The WTA dropout technique can be easily applied in deep neural networks without incurring additional training variables. More importantly, DASNet can be seamlessly integrated with other compression techniques, such as weight pruning and quantization, without compromising accuracy. Our experiments on various networks and datasets present significant runtime speedups with negligible accuracy losses.

Enabling On-Device CNN Training by Self-Supervised Instance Filtering and Error Map Pruning

This work aims to enable on-device training of convolutional neural networks (CNNs) by reducing the computation cost at training time. CNN models are usually trained on high-performance computers and only the trained models are deployed to edge devices. But the statically trained model cannot adapt dynamically in a real environment and may result in low accuracy for new inputs. On-device training by learning from the real-world data after deployment can greatly improve accuracy. However, the high computation cost makes training prohibitive for resource-constrained devices. To tackle this problem, we explore the computational redundancies in training and reduce the computation cost by two complementary approaches: 1) self-supervised early instance filtering on data level and 2) error map pruning (EMP) on the algorithm level. The early instance filter selects important instances from the input stream to train the network and drops trivial ones. The EMP further prunes out insignificant computations when training with the selected instances. Extensive experiments show that the computation and energy cost is substantially reduced without any or with marginal accuracy loss. For example, when training ResNet-110 on CIFAR-10, we achieve 67.8% computation saving while preserving full accuracy and 75.1% computation saving with a marginal accuracy loss of 1.3%. When training LeNet on MNIST, we save 79% computation while boosting accuracy by 0.2%. Besides, practical energy saving is measured on edge platforms. We achieve 67.6% energy saving when training ResNet-110 on mobile GPU and 74.1% energy saving when training LeNet on MCU without any accuracy loss.

Superpixel and multi-atlas based fusion entropic model for the segmentation of X-ray images.

X-ray image segmentation is an important and crucial step for three-dimensional (3D) bone reconstruction whose final goal remains to increase effectiveness of computer-aided diagnosis, surgery and treatment plannings. However, this segmentation task is rather challenging, particularly when dealing with complicated human structures in the lower limb such as the patella, talus and pelvis. In this work, we present a multi-atlas fusion framework for the automatic segmentation of these complex bone regions from a single X-ray view. The first originality of the proposed approach lies in the use of a (training) dataset of co-registered/pre-segmented X-ray images of these aforementioned bone regions (or multi-atlas) to estimate a collection of superpixels allowing us to take into account all the nonlinear and local variability of bone regions existing in the training dataset and also to simplify the superpixel map pruning process related to our strategy. The second originality is to introduce a novel label propagation step based on the entropy concept for refining the resulting segmentation map into the most likely internal regions to the final consensus segmentation. In this framework, a leave-one-out cross-validation process was performed on 31 manually segmented radiographic image dataset for each bone structure in order to rigorously evaluate the efficiency of the proposed method. The proposed method resulted in more accurate segmentations compared to the probabilistic patch-based label fusion model (PB) and the classical patch-based majority voting fusion scheme (MV) using different registration strategies. Comparison with manual (gold standard) segmentations revealed that the good classification accuracy of our unsupervised segmentation scheme is, respectively, 93.79% for the patella, 88.3% for the talus and 85.02% for the pelvis; a score that falls within the range of accuracy levels of manual segmentations (due to the intra inter/observer variability).

An extended depth-first search algorithm for optimal triangulation of Bayesian networks

The junction tree algorithm is currently the most popular algorithm for exact inference on Bayesian networks. To improve the time complexity of the junction tree algorithm, we need to find a triangulation with the optimal total table size. For this purpose, Ottosen and Vomlel have proposed a depth-first search (DFS) algorithm. They also introduced several techniques to improve the DFS algorithm, including dynamic clique maintenance and coalescing map pruning. Nevertheless, the efficiency and scalability of that algorithm leave much room for improvement. First, the dynamic clique maintenance allows to recompute some cliques. Second, in the worst case, the DFS algorithm explores the search space of all elimination orders, which has size n!, where n is the number of variables in the Bayesian network. To mitigate these problems, we propose an extended depth-first search (EDFS) algorithm. The new EDFS algorithm introduces the following two techniques as improvements to the DFS algorithm: (1) a new dynamic clique maintenance algorithm that computes only those cliques that contain a new edge, and (2) a new pruning rule, called pivot clique pruning. The new dynamic clique maintenance algorithm explores a smaller search space and runs faster than the Ottosen and Vomlel approach. This improvement can decrease the overhead cost of the DFS algorithm, and the pivot clique pruning reduces the size of the search space by a factor of O(n2). Our empirical results show that our proposed algorithm finds an optimal triangulation markedly faster than the state-of-the-art algorithm does.

Significance map pruning and other enhancements to SPIHT image coding algorithm

This paper proposes several enhancements to the Set Partitioning in Hierarchical Trees (SPIHT) image coding algorithm without changing the original algorithm's general skeleton. First and foremost, a method for significance map pruning based on a rate-distortion criterion is introduced. Specifically, the (Type A) sets of wavelet coefficients with small ratios of estimated distortion reduction to estimated rate contribution are deemed insignificant and effectively pruned. Even though determining such sets requires the computational complexity of the encoder to increase considerably with respect to the original SPIHT encoder, the original SPIHT decoder may still be used to decode the generated bitstream with a low computational complexity. The paper also proposes three low complexity enhancements by more sophisticated use of the adaptive arithmetic coder. Simulation results demonstrate that all these enhancements yield modest compression gains at moderate to high rates.