We introduce a discriminative multimodal descriptor based on a pair of sensor readings: a point cloud from a LiDAR and an image from an RGB camera. Our descriptor, named MinkLoc++, can be used for place recognition, re-localization and loop closure purposes in robotics or autonomous vehicles applications. We use late fusion approach, where each modality is processed separately and fused in the final part of the processing pipeline. The proposed method achieves state-of-the-art performance on standard place recognition benchmarks. We also identify dominating modality problem when training a multimodal descriptor. The problem manifests itself when the network focuses on a modality with a larger overfit to the training data. This drives the loss down during the training but leads to suboptimal performance on the evaluation set. In this work we describe how to detect and mitigate such risk when using a deep metric learning approach to train a multimodal neural network. Our code is publicly available on the project website: https://github.com/jac99/MinkLoc3DRGB.
Semi-structured query systems for document-oriented databases have many real applications. One particular application that we are interested in is matching each financial receipt image with its corresponding place of interest (POI, e.g., restaurant) in the nationwide database. The problem is especially challenging in the real production environment where many similar or incomplete entries exist in the database and queries are noisy (e.g., errors in optical character recognition). In this work, we aim to address practical challenges when using embedding-based retrieval for the query grounding problem in semi-structured data. Leveraging recent advancements in deep language encoding for retrieval, we conduct extensive experiments to find the most effective combination of modules for the embedding and retrieval of both query and database entries without any manually engineered component. The proposed model significantly outperforms the conventional manual pattern-based model while requiring much less development and maintenance cost. We also discuss some core observations in our experiments, which could be helpful for practitioners working on a similar problem in other domains.
One-pixel attack is a curious way of deceiving neural network classifier by changing only one pixel in the input image. The full potential and boundaries of this attack method are not yet fully understood. In this research, the successful and unsuccessful attacks are studied in more detail to illustrate the working mechanisms of a one-pixel attack. The data comes from our earlier studies where we applied the attack against medical imaging. We used a real breast cancer tissue dataset and a real classifier as the attack target. This research presents ways to analyze chromatic and spatial distributions of one-pixel attacks. In addition, we present one-pixel attack confidence maps to illustrate the behavior of the target classifier. We show that the more effective attacks change the color of the pixel more, and that the successful attacks are situated at the center of the images. This kind of analysis is not only useful for understanding the behavior of the attack but also the qualities of the classifying neural network.
Classical Visual Servoing (VS) rely on handcrafted visual features, which limit their generalizability. Recently, a number of approaches, some based on Deep Neural Networks, have been proposed to overcome this limitation by comparing directly the entire target and current camera images. However, by getting rid of the visual features altogether, those approaches require the target and current images to be essentially similar, which precludes the generalization to unknown, cluttered, scenes. Here we propose to perform VS based on visual features as in classical VS approaches but, contrary to the latter, we leverage recent breakthroughs in Deep Learning to automatically extract and match the visual features. By doing so, our approach enjoys the advantages from both worlds: (i) because our approach is based on visual features, it is able to steer the robot towards the object of interest even in presence of significant distraction in the background; (ii) because the features are automatically extracted and matched, our approach can easily and automatically generalize to unseen objects and scenes. In addition, we propose to use a render engine to synthesize the target image, which offers a further level of generalization. We demonstrate these advantages in a robotic grasping task, where the robot is able to steer, with high accuracy, towards the object to grasp, based simply on an image of the object rendered from the camera view corresponding to the desired robot grasping pose.
We present HyperMorph, a learning-based strategy for deformable image registration that removes the need to tune important registration hyperparameters during training. Classical registration methods solve an optimization problem to find a set of spatial correspondences between two images, while learning-based methods leverage a training dataset to learn a function that generates these correspondences. The quality of the results for both types of techniques depends greatly on the choice of hyperparameters. Unfortunately, hyperparameter tuning is time-consuming and typically involves training many separate models with various hyperparameter values, potentially leading to suboptimal results. To address this inefficiency, we introduce amortized hyperparameter learning for image registration, a novel strategy to learn the effects of hyperparameters on deformation fields. The proposed framework learns a hypernetwork that takes in an input hyperparameter and modulates a registration network to produce the optimal deformation field for that hyperparameter value. In effect, this strategy trains a single, rich model that enables rapid, fine-grained discovery of hyperparameter values from a continuous interval at test-time. We demonstrate that this approach can be used to optimize multiple hyperparameters considerably faster than existing search strategies, leading to a reduced computational and human burden and increased flexibility. We also show that this has several important benefits, including increased robustness to initialization and the ability to rapidly identify optimal hyperparameter values specific to a registration task, dataset, or even a single anatomical region - all without retraining the HyperMorph model. Our code is publicly available at http://voxelmorph.mit.edu.
Compensating scarce measurements by inferring them from computational models is a way to address ill-posed inverse problems. We tackle Limited Angle Tomography by completing the set of acquisitions using a generative model and prior-knowledge about the scanned object. Using a Generative Adversarial Network as model and Computer-Assisted Design data as shape prior, we demonstrate a quantitative and qualitative advantage of our technique over other state-of-the-art methods. Inferring a substantial number of consecutive missing measurements, we offer an alternative to other image inpainting techniques that fall short of providing a satisfying answer to our research question: can X-Ray exposition be reduced by using generative models to infer lacking measurements?
Image recognition with prototypes is considered an interpretable alternative for black box deep learning models. Classification depends on the extent to which a test image "looks like" a prototype. However, perceptual similarity for humans can be different from the similarity learnt by the model. A user is unaware of the underlying classification strategy and does not know which image characteristics (e.g., color or shape) is the dominant characteristic for the decision. We address this ambiguity and argue that prototypes should be explained. Only visualizing prototypes can be insufficient for understanding what a prototype exactly represents, and why a prototype and an image are considered similar. We improve interpretability by automatically enhancing prototypes with extra information about visual characteristics considered important by the model. Specifically, our method quantifies the influence of color hue, shape, texture, contrast and saturation in a prototype. We apply our method to the existing Prototypical Part Network (ProtoPNet) and show that our explanations clarify the meaning of a prototype which might have been interpreted incorrectly otherwise. We also reveal that visually similar prototypes can have the same explanations, indicating redundancy. Because of the generality of our approach, it can improve the interpretability of any similarity-based method for prototypical image recognition.
Single image de-hazing is a challenging problem, and it is far from solved. Most current solutions require paired image datasets that include both hazy images and their corresponding haze-free ground-truth images. However, in reality, lighting conditions and other factors can produce a range of haze-free images that can serve as ground truth for a hazy image, and a single ground truth image cannot capture that range. This limits the scalability and practicality of paired image datasets in real-world applications. In this paper, we focus on unpaired single image de-hazing and we do not rely on the ground truth image or physical scattering model. We reduce the image de-hazing problem to an image-to-image translation problem and propose a dehazing Global-Local Cycle-consistent Generative Adversarial Network (Dehaze-GLCGAN). Generator network of Dehaze-GLCGAN combines an encoder-decoder architecture with residual blocks to better recover the haze free scene. We also employ a global-local discriminator structure to deal with spatially varying haze. Through ablation study, we demonstrate the effectiveness of different factors in the performance of the proposed network. Our extensive experiments over three benchmark datasets show that our network outperforms previous work in terms of PSNR and SSIM while being trained on smaller amount of data compared to other methods.
Deep neural networks are parameterised by weights that encode feature representations, whose performance is dictated through generalisation by using large-scale feature-rich datasets. The lack of large-scale labelled 3D medical imaging datasets restrict constructing such generalised networks. In this work, a novel 3D segmentation network, Fabric Image Representation Networks (FIRENet), is proposed to extract and encode generalisable feature representations from multiple medical image datasets in a large-scale manner. FIRENet learns image specific feature representations by way of 3D fabric network architecture that contains exponential number of sub-architectures to handle various protocols and coverage of anatomical regions and structures. The fabric network uses Atrous Spatial Pyramid Pooling (ASPP) extended to 3D to extract local and image-level features at a fine selection of scales. The fabric is constructed with weighted edges allowing the learnt features to dynamically adapt to the training data at an architecture level. Conditional padding modules, which are integrated into the network to reinsert voxels discarded by feature pooling, allow the network to inherently process different-size images at their original resolutions. FIRENet was trained for feature learning via automated semantic segmentation of pelvic structures and obtained a state-of-the-art median DSC score of 0.867. FIRENet was also simultaneously trained on MR (Magnatic Resonance) images acquired from 3D examinations of musculoskeletal elements in the (hip, knee, shoulder) joints and a public OAI knee dataset to perform automated segmentation of bone across anatomy. Transfer learning was used to show that the features learnt through the pelvic segmentation helped achieve improved mean DSC scores of 0.962, 0.963, 0.945 and 0.986 for automated segmentation of bone across datasets.
We introduce the visual acoustic matching task, in which an audio clip is transformed to sound like it was recorded in a target environment. Given an image of the target environment and a waveform for the source audio, the goal is to re-synthesize the audio to match the target room acoustics as suggested by its visible geometry and materials. To address this novel task, we propose a cross-modal transformer model that uses audio-visual attention to inject visual properties into the audio and generate realistic audio output. In addition, we devise a self-supervised training objective that can learn acoustic matching from in-the-wild Web videos, despite their lack of acoustically mismatched audio. We demonstrate that our approach successfully translates human speech to a variety of real-world environments depicted in images, outperforming both traditional acoustic matching and more heavily supervised baselines.