Event cameras provide a number of benefits over traditional cameras, such as the ability to track incredibly fast motions, high dynamic range, and low power consumption. However, their application into computer vision problems, many of which are primarily dominated by deep learning solutions, has been limited by the lack of labeled training data for events. In this work, we propose a method which leverages the existing labeled data for images by simulating events from a pair of temporal image frames, using a convolutional neural network. We train this network on pairs of images and events, using an adversarial discriminator loss and a pair of cycle consistency losses. The cycle consistency losses utilize a pair of pre-trained self-supervised networks which perform optical flow estimation and image reconstruction from events, and constrain our network to generate events which result in accurate outputs from both of these networks. Trained fully end to end, our network learns a generative model for events from images without the need for accurate modeling of the motion in the scene, exhibited by modeling based methods, while also implicitly modeling event noise. Using this simulator, we train a pair of downstream networks on object detection and 2D human pose estimation from events, using simulated data from large scale image datasets, and demonstrate the networks' abilities to generalize to datasets with real events.
Fine-tuning of large pre-trained image and language models on small customized datasets has become increasingly popular for improved prediction and efficient use of limited resources. Fine-tuning requires identification of best models to transfer-learn from and quantifying transferability prevents expensive re-training on all of the candidate models/tasks pairs. We show that the statistical problems with covariance estimation drive the poor performance of H-score [Bao et al., 2019] -- a common baseline for newer metrics -- and propose shrinkage-based estimator. This results in up to 80% absolute gain in H-score correlation performance, making it competitive with the state-of-the-art LogME measure by You et al. [2021]. Our shrinkage-based H-score is 3-55 times faster to compute compared to LogME. Additionally, we look into a less common setting of target (as opposed to source) task selection. We identify previously overlooked problems in such settings with different number of labels, class-imbalance ratios etc. for some recent metrics e.g., LEEP [Nguyen et al., 2020] that resulted in them being misrepresented as leading measures. We propose a correction and recommend measuring correlation performance against relative accuracy in such settings. We also outline the difficulties of comparing feature-dependent metrics, both supervised (e.g. H-score) and unsupervised measures (e.g., Maximum Mean Discrepancy [Long et al., 2015]), across source models/layers with different feature embedding dimension. We show that dimensionality reduction methods allow for meaningful comparison across models and improved performance of some of these measures. We investigate performance of 14 different supervised and unsupervised metrics and demonstrate that even unsupervised metrics can identify the leading models for domain adaptation. We support our findings with ~65,000 (fine-tuning trials) experiments.
We present a novel approach to robotic grasp planning using both a learned grasp proposal network and a learned 3D shape reconstruction network. Our system generates 6-DOF grasps from a single RGB-D image of the target object, which is provided as input to both networks. By using the geometric reconstruction to refine the the candidate grasp produced by the grasp proposal network, our system is able to accurately grasp both known and unknown objects, even when the grasp location on the object is not visible in the input image. This paper presents the network architectures, training procedures, and grasp refinement method that comprise our system. Hardware experiments demonstrate the efficacy of our system at grasping both known and unknown objects (91% success rate). We additionally perform ablation studies that show the benefits of combining a learned grasp proposal with geometric reconstruction for grasping, and also show that our system outperforms several baselines in a grasping task.
Systematic error, which is not determined by chance, often refers to the inaccuracy (involving either the observation or measurement process) inherent to a system. In this paper, we exhibit some long-neglected but frequent-happening adversarial examples caused by systematic error. More specifically, we find the trained neural network classifier can be fooled by inconsistent implementations of image decoding and resize. This tiny difference between these implementations often causes an accuracy drop from training to deployment. To benchmark these real-world adversarial examples, we propose ImageNet-S dataset, which enables researchers to measure a classifier's robustness to systematic error. For example, we find a normal ResNet-50 trained on ImageNet can have 1%-5% accuracy difference due to the systematic error. Together our evaluation and dataset may aid future work toward real-world robustness and practical generalization.
Application of underwater robots are on the rise, most of them are dependent on sonar for underwater vision, but the lack of strong perception capabilities limits them in this task. An important issue in sonar perception is matching image patches, which can enable other techniques like localization, change detection, and mapping. There is a rich literature for this problem in color images, but for acoustic images, it is lacking, due to the physics that produce these images. In this paper we improve on our previous results for this problem (Valdenegro-Toro et al, 2017), instead of modeling features manually, a Convolutional Neural Network (CNN) learns a similarity function and predicts if two input sonar images are similar or not. With the objective of improving the sonar image matching problem further, three state of the art CNN architectures are evaluated on the Marine Debris dataset, namely DenseNet, and VGG, with a siamese or two-channel architecture, and contrastive loss. To ensure a fair evaluation of each network, thorough hyper-parameter optimization is executed. We find that the best performing models are DenseNet Two-Channel network with 0.955 AUC, VGG-Siamese with contrastive loss at 0.949 AUC and DenseNet Siamese with 0.921 AUC. By ensembling the top performing DenseNet two-channel and DenseNet-Siamese models overall highest prediction accuracy obtained is 0.978 AUC, showing a large improvement over the 0.91 AUC in the state of the art.
This work presents a comparison of machine learning algorithms that are implemented to segment the characters of text presented as an image. The algorithms are designed to work on degraded documents with text that is not aligned in an organized fashion. The paper investigates the use of Support Vector Machines, K-Nearest Neighbor algorithm and an Encoder Network to perform the operation of character spotting. Character Spotting involves extracting potential characters from a stream of text by selecting regions bound by white space.
This paper proposes a novel Non-Local Attention Optimized Deep Image Compression (NLAIC) framework, which is built on top of the popular variational auto-encoder (VAE) structure. Our NLAIC framework embeds non-local operations in the encoders and decoders for both image and latent feature probability information (known as hyperprior) to capture both local and global correlations, and apply attention mechanism to generate masks that are used to weigh the features for the image and hyperprior, which implicitly adapt bit allocation for different features based on their importance. Furthermore, both hyperpriors and spatial-channel neighbors of the latent features are used to improve entropy coding. The proposed model outperforms the existing methods on Kodak dataset, including learned (e.g., Balle2019, Balle2018) and conventional (e.g., BPG, JPEG2000, JPEG) image compression methods, for both PSNR and MS-SSIM distortion metrics.
This is the report for the PRIM project in Telecom Paris. This report is about applications based on spatial-frequency transform and deep learning techniques. In this report, there are two main works. The first work is about the enhanced JPEG compression method based on deep learning. we propose a novel method to highly enhance the JPEG compression by transmitting fewer image data at the sender's end. At the receiver's end, we propose a DC recovery algorithm together with the deep residual learning framework to recover images with high quality. The second work is about adversarial examples defenses based on signal processing. We propose the wavelet extension method to extend image data features, which makes it more difficult to generate adversarial examples. We further adopt wavelet denoising to reduce the influence of the adversarial perturbations. With intensive experiments, we demonstrate that both works are effective in their application scenarios.
Capsule networks (CapsNets) are an emerging trend in image processing. In contrast to a convolutional neural network, CapsNets are not vulnerable to object deformation, as the relative spatial information of the objects is preserved across the network. However, their complexity is mainly related with the capsule structure and the dynamic routing mechanism, which makes it almost unreasonable to deploy a CapsNet, in its original form, in a resource-constrained device powered by a small microcontroller (MCU). In an era where intelligence is rapidly shifting from the cloud to the edge, this high complexity imposes serious challenges to the adoption of CapsNets at the very edge. To tackle this issue, we present an API for the execution of quantized CapsNets in Cortex-M and RISC-V MCUs. Our software kernels extend the Arm CMSIS-NN and RISC-V PULP-NN, to support capsule operations with 8-bit integers as operands. Along with it, we propose a framework to perform post training quantization of a CapsNet. Results show a reduction in memory footprint of almost 75%, with a maximum accuracy loss of 1%. In terms of throughput, our software kernels for the Arm Cortex-M are, at least, 5.70x faster than a pre-quantized CapsNet running on an NVIDIA GTX 980 Ti graphics card. For RISC-V, the throughout gain increases to 26.28x and 56.91x for a single- and octa-core configuration, respectively.
Real-world object sampling produces long-tailed distributions requiring exponentially more images for rare types. Zero-shot detection, which aims to detect unseen objects, is one direction to address this problem. A dataset such as COCO is extensively annotated across many images but with a sparse number of categories and annotating all object classes across a diverse domain is expensive and challenging. To advance zero-shot detection, we develop a Vision-Language distillation method that aligns both image and text embeddings from a zero-shot pre-trained model such as CLIP to a modified semantic prediction head from a one-stage detector like YOLOv5. With this method, we are able to train an object detector that achieves state-of-the-art accuracy on the COCO zero-shot detection splits with fewer model parameters. During inference, our model can be adapted to detect any number of object classes without additional training. We also find that the improvements provided by the scaling of our method are consistent across various YOLOv5 scales. Furthermore, we develop a self-training method that provides a significant score improvement without needing extra images nor labels.