Convolutional neural network (CNN) inference using fully homomorphic encryption (FHE) is a promising private inference (PI) solution due to the capability of FHE that enables offloading the whole computation process to the server while protecting the privacy of sensitive user data. However, prior FHEbased CNN (HCNN) implementations are far from being practical due to the high computational and memory overheads of FHE. To overcome this limitation, we present HyPHEN, a deep HCNN construction that features an efficient FHE convolution algorithm, data packing methods (hybrid packing and image slicing), and FHE-specific optimizations. Such enhancements enable HyPHEN to substantially reduce the memory footprint and the number of expensive homomorphic operations, such as ciphertext rotation and bootstrapping. As a result, HyPHEN brings the latency of HCNN CIFAR-10 inference down to a practical level at 1.40s (ResNet20) and demonstrates HCNN ImageNet inference for the first time at 16.87s (ResNet18).
3D visual perception tasks based on multi-camera images are essential for autonomous driving systems. Latest work in this field performs 3D object detection by leveraging multi-view images as an input and iteratively enhancing object queries (object proposals) by cross-attending multi-view features. However, individual backbone features are not updated with multi-view features and it stays as a mere collection of the output of the single-image backbone network. Therefore we propose 3M3D: A Multi-view, Multi-path, Multi-representation for 3D Object Detection where we update both multi-view features and query features to enhance the representation of the scene in both fine panoramic view and coarse global view. Firstly, we update multi-view features by multi-view axis self-attention. It will incorporate panoramic information in the multi-view features and enhance understanding of the global scene. Secondly, we update multi-view features by self-attention of the ROI (Region of Interest) windows which encodes local finer details in the features. It will help exchange the information not only along the multi-view axis but also along the other spatial dimension. Lastly, we leverage the fact of multi-representation of queries in different domains to further boost the performance. Here we use sparse floating queries along with dense BEV (Bird's Eye View) queries, which are later post-processed to filter duplicate detections. Moreover, we show performance improvements on nuScenes benchmark dataset on top of our baselines.
Video understanding tasks take many forms, from action detection to visual query localization and spatio-temporal grounding of sentences. These tasks differ in the type of inputs (only video, or video-query pair where query is an image region or sentence) and outputs (temporal segments or spatio-temporal tubes). However, at their core they require the same fundamental understanding of the video, i.e., the actors and objects in it, their actions and interactions. So far these tasks have been tackled in isolation with individual, highly specialized architectures, which do not exploit the interplay between tasks. In contrast, in this paper, we present a single, unified model for tackling query-based video understanding in long-form videos. In particular, our model can address all three tasks of the Ego4D Episodic Memory benchmark which entail queries of three different forms: given an egocentric video and a visual, textual or activity query, the goal is to determine when and where the answer can be seen within the video. Our model design is inspired by recent query-based approaches to spatio-temporal grounding, and contains modality-specific query encoders and task-specific sliding window inference that allow multi-task training with diverse input modalities and different structured outputs. We exhaustively analyze relationships among the tasks and illustrate that cross-task learning leads to improved performance on each individual task, as well as the ability to generalize to unseen tasks, such as zero-shot spatial localization of language queries.
Ever since their conception, Transformers have taken over traditional sequence models in many tasks, such as NLP, image classification, and video/audio processing, for their fast training and superior performance. Much of the merit is attributable to positional encoding and multi-head attention. However, Transformers fall short in learning long-range dependencies mainly due to the quadratic complexity scaled with context length, in terms of both time and space. Consequently, over the past five years, a myriad of methods has been proposed to make Transformers more efficient. In this work, we first take a step back, study and compare existing solutions to long-sequence modeling in terms of their pure mathematical formulation. Specifically, we summarize them using a unified template, given their shared nature of token mixing. Through benchmarks, we then demonstrate that long context length does yield better performance, albeit application-dependent, and traditional Transformer models fall short in taking advantage of long-range dependencies. Next, inspired by emerging sparse models of huge capacity, we propose a machine learning system for handling million-scale dependencies. As a proof of concept, we evaluate the performance of one essential component of this system, namely, the distributed multi-head attention. We show that our algorithm can scale up attention computation by almost $40\times$ using four GeForce RTX 4090 GPUs, compared to vanilla multi-head attention mechanism. We believe this study is an instrumental step towards modeling million-scale dependencies.
In this paper, we introduce Planet-CR, a benchmark dataset for high-resolution cloud removal with multi-modal and multi-resolution data fusion. Planet-CR is the first public dataset for cloud removal to feature globally sampled high resolution optical observations, in combination with paired radar measurements as well as pixel-level land cover annotations. It provides solid basis for exhaustive evaluation in terms of generating visually pleasing textures and semantically meaningful structures. With this dataset, we consider the problem of cloud removal in high resolution optical remote sensing imagery by integrating multi-modal and multi-resolution information. Existing multi-modal data fusion based methods, which assume the image pairs are aligned pixel-to-pixel, are hence not appropriate for this problem. To this end, we design a new baseline named Align-CR to perform the low-resolution SAR image guided high-resolution optical image cloud removal. It implicitly aligns the multi-modal and multi-resolution data during the reconstruction process to promote the cloud removal performance. The experimental results demonstrate that the proposed Align-CR method gives the best performance in both visual recovery quality and semantic recovery quality. The project is available at https://github.com/zhu-xlab/Planet-CR, and hope this will inspire future research.
Suppression of thoracic bone shadows on chest X-rays (CXRs) has been indicated to improve the diagnosis of pulmonary disease. Previous approaches can be categorized as unsupervised physical and supervised deep learning models. Nevertheless, with physical models able to preserve morphological details but at the cost of extremely long processing time, existing DL methods face challenges of gathering sufficient/qualitative ground truth (GT) for robust training, thus leading to failure in maintaining clinically acceptable false positive rates. We hereby propose a generalizable yet efficient workflow of two stages: (1) training pairs generation with GT bone shadows eliminated in by a physical model in spatially transformed gradient fields. (2) fully supervised image denoising network training on stage-one datasets for fast rib removal on incoming CXRs. For step two, we designed a densely connected network called SADXNet, combined with peak signal to noise ratio and multi-scale structure similarity index measure objective minimization to suppress bony structures. The SADXNet organizes spatial filters in U shape (e.g., X=7; filters = 16, 64, 256, 512, 256, 64, 16) and preserves the feature map dimension throughout the network flow. Visually, SADXNet can suppress the rib edge and that near the lung wall/vertebra without jeopardizing the vessel/abnormality conspicuity. Quantitively, it achieves RMSE of ~0 during testing with one prediction taking <1s. Downstream tasks including lung nodule detection as well as common lung disease classification and localization are used to evaluate our proposed rib suppression mechanism. We observed 3.23% and 6.62% area under the curve (AUC) increase as well as 203 and 385 absolute false positive decrease for lung nodule detection and common lung disease localization, separately.
In order to get raw images of high quality for downstream Image Signal Process (ISP), in this paper we present an Efficient Locally Multiplicative Transformer called ELMformer for raw image restoration. ELMformer contains two core designs especially for raw images whose primitive attribute is single-channel. The first design is a Bi-directional Fusion Projection (BFP) module, where we consider both the color characteristics of raw images and spatial structure of single-channel. The second one is that we propose a Locally Multiplicative Self-Attention (L-MSA) scheme to effectively deliver information from the local space to relevant parts. ELMformer can efficiently reduce the computational consumption and perform well on raw image restoration tasks. Enhanced by these two core designs, ELMformer achieves the highest performance and keeps the lowest FLOPs on raw denoising and raw deblurring benchmarks compared with state-of-the-arts. Extensive experiments demonstrate the superiority and generalization ability of ELMformer. On SIDD benchmark, our method has even better denoising performance than ISP-based methods which need huge amount of additional sRGB training images. The codes are release at https://github.com/leonmakise/ELMformer.
Lacking supervised data is an issue while training deep neural networks (DNNs), mainly when considering medical and biological data where supervision is expensive. Recently, Embedded Pseudo-Labeling (EPL) addressed this problem by using a non-linear projection (t-SNE) from a feature space of the DNN to a 2D space, followed by semi-supervised label propagation using a connectivity-based method (OPFSemi). We argue that the performance of the final classifier depends on the data separation present in the latent space and visual separation present in the projection. We address this by first proposing to use contrastive learning to produce the latent space for EPL by two methods (SimCLR and SupCon) and by their combination, and secondly by showing, via an extensive set of experiments, the aforementioned correlations between data separation, visual separation, and classifier performance. We demonstrate our results by the classification of five real-world challenging image datasets of human intestinal parasites with only 1% supervised samples.
As participants of the MediaEval 2022 Sport Task, we propose a two-stream network approach for the classification and detection of table tennis strokes. Each stream is a succession of 3D Convolutional Neural Network (CNN) blocks using attention mechanisms. Each stream processes different 4D inputs. Our method utilizes raw RGB data and pose information computed from MMPose toolbox. The pose information is treated as an image by applying the pose either on a black background or on the original RGB frame it has been computed from. Best performance is obtained by feeding raw RGB data to one stream, Pose + RGB (PRGB) information to the other stream and applying late fusion on the features. The approaches were evaluated on the provided TTStroke-21 data sets. We can report an improvement in stroke classification, reaching 87.3% of accuracy, while the detection does not outperform the baseline but still reaches an IoU of 0.349 and mAP of 0.110.
Large vision-language models are generally applicable to many downstream tasks, but come at an exorbitant training cost that only large institutions can afford. This paper trades generality for efficiency and presents Curation in Training (CiT), a simple and efficient vision-text learning algorithm that couples a data objective into training. CiT automatically yields quality data to speed-up contrastive image-text training and alleviates the need for an offline data filtering pipeline, allowing broad data sources (including raw image-text pairs from the web). CiT contains two loops: an outer loop curating the training data and an inner loop consuming the curated training data. The text encoder connects the two loops. Given metadata for tasks of interest, e.g., class names, and a large pool of image-text pairs, CiT alternatively selects relevant training data from the pool by measuring the similarity of their text embeddings and embeddings of the metadata. In our experiments, we observe that CiT can speed up training by over an order of magnitude, especially if the raw data size is large.