Absence of large-scale labeled data in the practitioner's target domain can be a bottleneck to applying machine learning algorithms in practice. Transfer learning is a popular strategy for leveraging additional data to improve the downstream performance, but finding the most relevant data to transfer from can be challenging. Neural Data Server (NDS), a search engine that recommends relevant data for a given downstream task, has been previously proposed to address this problem. NDS uses a mixture of experts trained on data sources to estimate similarity between each source and the downstream task. Thus, the computational cost to each user grows with the number of sources. To address these issues, we propose Scalable Neural Data Server (SNDS), a large-scale search engine that can theoretically index thousands of datasets to serve relevant ML data to end users. SNDS trains the mixture of experts on intermediary datasets during initialization, and represents both data sources and downstream tasks by their proximity to the intermediary datasets. As such, computational cost incurred by SNDS users remains fixed as new datasets are added to the server. We validate SNDS on a plethora of real world tasks and find that data recommended by SNDS improves downstream task performance over baselines. We also demonstrate the scalability of SNDS by showing its ability to select relevant data for transfer outside of the natural image setting.
Deep convolutional networks have attracted great attention in image restoration and enhancement. Generally, restoration quality has been improved by building more and more convolutional block. However, these methods mostly learn a specific model to handle all images and ignore difficulty diversity. In other words, an area in the image with high frequency tend to lose more information during compressing while an area with low frequency tends to lose less. In this article, we adrress the efficiency issue in image SR by incorporating a patch-wise rolling network(PRN) to content-adaptively recover images according to difficulty levels. In contrast to existing studies that ignore difficulty diversity, we adopt different stage of a neural network to perform image restoration. In addition, we propose a rolling strategy that utilizes the parameters of each stage more flexible. Extensive experiments demonstrate that our model not only shows a significant acceleration but also maintain state-of-the-art performance.
Recently, deep learning methods have shown great success in 3D point cloud upsampling. Among these methods, many feature expansion units were proposed to complete point expansion at the end. In this paper, we compare various feature expansion units by both theoretical analysis and quantitative experiments. We show that most of the existing feature expansion units process each point feature independently, while ignoring the feature interaction among different points. Further, inspired by upsampling module of image super-resolution and recent success of dynamic graph CNN on point clouds, we propose a novel feature expansion units named ProEdgeShuffle. Experiments show that our proposed method can achieve considerable improvement over previous feature expansion units.
We propose a method to infer a dense depth map from a single image, its calibration, and the associated sparse point cloud. In order to leverage existing models that produce putative depth maps (teacher models), we propose an adaptive knowledge distillation approach that yields a positive congruent training process, where a student model avoids learning the error modes of the teachers. We consider the scenario of a blind ensemble where we do not have access to ground truth for model selection nor training. The crux of our method, termed Monitored Distillation, lies in a validation criterion that allows us to learn from teachers by choosing predictions that best minimize the photometric reprojection error for a given image. The result of which is a distilled depth map and a confidence map, or "monitor", for how well a prediction from a particular teacher fits the observed image. The monitor adaptively weights the distilled depth where, if all of the teachers exhibit high residuals, the standard unsupervised image reconstruction loss takes over as the supervisory signal. On indoor scenes (VOID), we outperform blind ensembling baselines by 13.3% and unsupervised methods by 20.3%; we boast a 79% model size reduction while maintaining comparable performance to the best supervised method. For outdoors (KITTI), we tie for 5th overall on the benchmark despite not using ground truth.
Visual question answering (VQA) in surgery is largely unexplored. Expert surgeons are scarce and are often overloaded with clinical and academic workloads. This overload often limits their time answering questionnaires from patients, medical students or junior residents related to surgical procedures. At times, students and junior residents also refrain from asking too many questions during classes to reduce disruption. While computer-aided simulators and recording of past surgical procedures have been made available for them to observe and improve their skills, they still hugely rely on medical experts to answer their questions. Having a Surgical-VQA system as a reliable 'second opinion' could act as a backup and ease the load on the medical experts in answering these questions. The lack of annotated medical data and the presence of domain-specific terms has limited the exploration of VQA for surgical procedures. In this work, we design a Surgical-VQA task that answers questionnaires on surgical procedures based on the surgical scene. Extending the MICCAI endoscopic vision challenge 2018 dataset and workflow recognition dataset further, we introduce two Surgical-VQA datasets with classification and sentence-based answers. To perform Surgical-VQA, we employ vision-text transformers models. We further introduce a residual MLP-based VisualBert encoder model that enforces interaction between visual and text tokens, improving performance in classification-based answering. Furthermore, we study the influence of the number of input image patches and temporal visual features on the model performance in both classification and sentence-based answering.
Towards a safe and comfortable driving, road scene segmentation is a rudimentary problem in camera-based advance driver assistance systems (ADAS). Despite of the great achievement of Convolutional Neural Networks (CNN) for semantic segmentation task, the high computational efforts of CNN based methods is still a challenging area. In recent work, we proposed a novel approach to utilise the advantages of CNNs for the task of road segmentation at reasonable computational effort. The runtime benefits from using irregular super pixels as basis for the input for the CNN rather than the image grid, which tremendously reduces the input size. Although, this method achieved remarkable low computational time in both training and testing phases, the lower resolution of the super pixel domain yields naturally lower accuracy compared to high cost state of the art methods. In this work, we focus on a refinement of the road segmentation utilising a Conditional Random Field (CRF).The refinement procedure is limited to the super pixels touching the predicted road boundary to keep the additional computational effort low. Reducing the input to the super pixel domain allows the CNNs structure to stay small and efficient to compute while keeping the advantage of convolutional layers and makes them eligible for ADAS. Applying CRF compensate the trade off between accuracy and computational efficiency. The proposed system obtained comparable performance among the top performing algorithms on the KITTI road benchmark and its fast inference makes it particularly suitable for realtime applications.
Automatic colourisation of grey-scale images is the process of creating a full-colour image from the grey-scale prior. It is an ill-posed problem, as there are many plausible colourisations for a given grey-scale prior. The current SOTA in auto-colourisation involves image-to-image type Deep Convolutional Neural Networks with Generative Adversarial Networks showing the greatest promise. The end goal of colourisation is to produce full colour images that appear plausible to the human viewer, but human assessment is costly and time consuming. This work assesses how well commonly used objective measures correlate with human opinion. We also attempt to determine what facets of colourisation have the most significant effect on human opinion. For each of 20 images from the BSD dataset, we create 65 recolourisations made up of local and global changes. Opinion scores are then crowd sourced using the Amazon Mechanical Turk and together with the images this forms an extensible dataset called the Human Evaluated Colourisation Dataset (HECD). While we find statistically significant correlations between human-opinion scores and a small number of objective measures, the strength of the correlations is low. There is also evidence that human observers are most intolerant to an incorrect hue of naturally occurring objects.
We, as human beings, can understand and picture a familiar scene from arbitrary viewpoints given a single image, whereas this is still a grand challenge for computers. We hereby present a novel solution to mimic such human perception capability based on a new paradigm of amodal 3D scene understanding with neural rendering for a closed scene. Specifically, we first learn the prior knowledge of the objects in a closed scene via an offline stage, which facilitates an online stage to understand the room with unseen furniture arrangement. During the online stage, given a panoramic image of the scene in different layouts, we utilize a holistic neural-rendering-based optimization framework to efficiently estimate the correct 3D scene layout and deliver realistic free-viewpoint rendering. In order to handle the domain gap between the offline and online stage, our method exploits compositional neural rendering techniques for data augmentation in the offline training. The experiments on both synthetic and real datasets demonstrate that our two-stage design achieves robust 3D scene understanding and outperforms competing methods by a large margin, and we also show that our realistic free-viewpoint rendering enables various applications, including scene touring and editing. Code and data are available on the project webpage: https://zju3dv.github.io/nr_in_a_room/.
Recent studies have shown that many deep metric learning loss functions perform very similarly under the same experimental conditions. One potential reason for this unexpected result is that all losses let the network focus on similar image regions or properties. In this paper, we investigate this by conducting a two-step analysis to extract and compare the learned visual features of the same model architecture trained with different loss functions: First, we compare the learned features on the pixel level by correlating saliency maps of the same input images. Second, we compare the clustering of embeddings for several image properties, e.g. object color or illumination. To provide independent control over these properties, photo-realistic 3D car renders similar to images in the Cars196 dataset are generated. In our analysis, we compare 14 pretrained models from a recent study and find that, even though all models perform similarly, different loss functions can guide the model to learn different features. We especially find differences between classification and ranking based losses. Our analysis also shows that some seemingly irrelevant properties can have significant influence on the resulting embedding. We encourage researchers from the deep metric learning community to use our methods to get insights into the features learned by their proposed methods.
In recent years , there has been an upsurge in a new form of entertainment medium called memes. These memes although seemingly innocuous have transcended onto the boundary of online harassment against women and created an unwanted bias against them . To help alleviate this problem , we propose an early fusion model for prediction and identification of misogynistic memes and its type in this paper for which we participated in SemEval-2022 Task 5 . The model receives as input meme image with its text transcription with a target vector. Given that a key challenge with this task is the combination of different modalities to predict misogyny, our model relies on pretrained contextual representations from different state-of-the-art transformer-based language models and pretrained image pretrained models to get an effective image representation. Our model achieved competitive results on both SubTask-A and SubTask-B with the other competition teams and significantly outperforms the baselines.