Recovering high quality surfaces from noisy point clouds, known as point cloud denoising, is a fundamental yet challenging problem in geometry processing. Most of the existing methods either directly denoise the noisy input or filter raw normals followed by updating point positions. Motivated by the essential interplay between point cloud denoising and normal filtering, we revisit point cloud denoising from a multitask perspective, and propose an end-to-end network, named PCDNF, to denoise point clouds via joint normal filtering. In particular, we introduce an auxiliary normal filtering task to help the overall network remove noise more effectively while preserving geometric features more accurately. In addition to the overall architecture, our network has two novel modules. On one hand, to improve noise removal performance, we design a shape-aware selector to construct the latent tangent space representation of the specific point by comprehensively considering the learned point and normal features and geometry priors. On the other hand, point features are more suitable for describing geometric details, and normal features are more conducive for representing geometric structures (e.g., sharp edges and corners). Combining point and normal features allows us to overcome their weaknesses. Thus, we design a feature refinement module to fuse point and normal features for better recovering geometric information. Extensive evaluations, comparisons, and ablation studies demonstrate that the proposed method outperforms state-of-the-arts for both point cloud denoising and normal filtering.
Generalized text representations are the foundation of many natural language understanding tasks. To fully utilize the different corpus, it is inevitable that models need to understand the relevance among them. However, many methods ignore the relevance and adopt a single-channel model (a coarse paradigm) directly for all tasks, which lacks enough rationality and interpretation. In addition, some existing works learn downstream tasks by stitches skill block(a fine paradigm), which might cause irrationalresults due to its redundancy and noise. Inthis work, we first analyze the task correlation through three different perspectives, i.e., data property, manual design, and model-based relevance, based on which the similar tasks are grouped together. Then, we propose a hierarchical framework with a coarse-to-fine paradigm, with the bottom level shared to all the tasks, the mid-level divided to different groups, and the top-level assigned to each of the tasks. This allows our model to learn basic language properties from all tasks, boost performance on relevant tasks, and reduce the negative impact from irrelevant tasks. Our experiments on 13 benchmark datasets across five natural language understanding tasks demonstrate the superiority of our method.
Video frame interpolation~(VFI) algorithms have improved considerably in recent years due to unprecedented progress in both data-driven algorithms and their implementations. Recent research has introduced advanced motion estimation or novel warping methods as the means to address challenging VFI scenarios. However, none of the published VFI works considers the spatially non-uniform characteristics of the interpolation error (IE). This work introduces such a solution. By closely examining the correlation between optical flow and IE, the paper proposes novel error prediction metrics that partition the middle frame into distinct regions corresponding to different IE levels. Building upon this IE-driven segmentation, and through the use of novel error-controlled loss functions, it introduces an ensemble of spatially adaptive interpolation units that progressively processes and integrates the segmented regions. This spatial ensemble results in an effective and computationally attractive VFI solution. Extensive experimentation on popular video interpolation benchmarks indicates that the proposed solution outperforms the current state-of-the-art (SOTA) in applications of current interest.
Deep recommender systems jointly leverage the retrieval and ranking operations to generate the recommendation result. The retriever targets selecting a small set of relevant candidates from the entire items with high efficiency; while the ranker, usually more precise but time-consuming, is supposed to identify the best items out of the retrieved candidates with high precision. However, the retriever and ranker are usually trained in poorly-cooperative ways, leading to limited recommendation performances when working as an entirety. In this work, we propose a novel DRS training framework CoRR(short for Cooperative Retriever and Ranker), where the retriever and ranker can be mutually reinforced. On one hand, the retriever is learned from recommendation data and the ranker via knowledge distillation; knowing that the ranker is more precise, the knowledge distillation may provide extra weak-supervision signals for the improvement of retrieval quality. On the other hand, the ranker is trained by learning to discriminate the truth positive items from hard negative candidates sampled from the retriever. With the iteration going on, the ranker may become more precise, which in return gives rise to informative training signals for the retriever; meanwhile, with the improvement of retriever, harder negative candidates can be sampled, which contributes to a higher discriminative capability of the ranker. To facilitate the effective conduct of CoRR, an asymptotic-unbiased approximation of KL divergence is introduced for the knowledge distillation over sampled items; besides, a scalable and adaptive strategy is developed to efficiently sample from the retriever. Comprehensive experimental studies are performed over four large-scale benchmark datasets, where CoRR improves the overall recommendation quality resulting from the cooperation between retriever and ranker.
Current state-of-the-art document retrieval solutions mainly follow an index-retrieve paradigm, where the index is hard to be optimized for the final retrieval target. In this paper, we aim to show that an end-to-end deep neural network unifying training and indexing stages can significantly improve the recall performance of traditional methods. To this end, we propose Neural Corpus Indexer (NCI), a sequence-to-sequence network that generates relevant document identifiers directly for a designated query. To optimize the recall performance of NCI, we invent a prefix-aware weight-adaptive decoder architecture, and leverage tailored techniques including query generation, semantic document identifiers and consistency-based regularization. Empirical studies demonstrated the superiority of NCI on a commonly used academic benchmark, achieving +51.9% relative improvement on NQ320k dataset compared to the best baseline.
Object detection is a difficult downstream task in computer vision. For the on-board edge computing platforms, a giant model is difficult to achieve the real-time detection requirement. And, a lightweight model built from a large number of the depth-wise separable convolutional layers cannot achieve the sufficient accuracy. We introduce a new method, GSConv, to lighten the model but maintain the accuracy. The GSConv balances the model's accuracy and speed better. And, we provide a design paradigm, slim-neck, to achieve a higher computational cost-effectiveness of the detectors. In experiments, our method obtains state-of-the-art results (e.g. 70.9% mAP0.5 for the SO-DA10M at a speed of ~100FPS on a Tesla T4) compared with the original networks. Code will be open source.
Pre-trained models have demonstrated superior power on many important tasks. However, it is still an open problem of designing effective pre-training strategies so as to promote the models' usability on dense retrieval. In this paper, we propose a novel pre-training framework for dense retrieval based on the Masked Auto-Encoder, known as RetroMAE. Our proposed framework is highlighted for the following critical designs: 1) a MAE based pre-training workflow, where the input sentence is polluted on both encoder and decoder side with different masks, and original sentence is reconstructed based on both sentence embedding and masked sentence; 2) asymmetric model architectures, with a large-scale expressive transformer for sentence encoding and a extremely simplified transformer for sentence reconstruction; 3) asymmetric masking ratios, with a moderate masking on the encoder side (15%) and an aggressive masking ratio on the decoder side (50~90%). We pre-train a BERT like encoder on English Wikipedia and BookCorpus, where it notably outperforms the existing pre-trained models on a wide range of dense retrieval benchmarks, like MS MARCO, Open-domain Question Answering, and BEIR.
A large-scale recommender system usually consists of recall and ranking modules. The goal of ranking modules (aka rankers) is to elaborately discriminate users' preference on item candidates proposed by recall modules. With the success of deep learning techniques in various domains, we have witnessed the mainstream rankers evolve from traditional models to deep neural models. However, the way that we design and use rankers remains unchanged: offline training the model, freezing the parameters, and deploying it for online serving. Actually, the candidate items are determined by specific user requests, in which underlying distributions (e.g., the proportion of items for different categories, the proportion of popular or new items) are highly different from one another in a production environment. The classical parameter-frozen inference manner cannot adapt to dynamic serving circumstances, making rankers' performance compromised. In this paper, we propose a new training and inference paradigm, termed as Ada-Ranker, to address the challenges of dynamic online serving. Instead of using parameter-frozen models for universal serving, Ada-Ranker can adaptively modulate parameters of a ranker according to the data distribution of the current group of item candidates. We first extract distribution patterns from the item candidates. Then, we modulate the ranker by the patterns to make the ranker adapt to the current data distribution. Finally, we use the revised ranker to score the candidate list. In this way, we empower the ranker with the capacity of adapting from a global model to a local model which better handles the current task.
For aquaculture resource evaluation and ecological environment monitoring, automatic detection and identification of marine organisms is critical. However, due to the low quality of underwater images and the characteristics of underwater biological, a lack of abundant features may impede traditional hand-designed feature extraction approaches or CNN-based object detection algorithms, particularly in complex underwater environment. Therefore, the goal of this paper is to perform object detection in the underwater environment. This paper proposed a novel method for capturing feature information, which adds the convolutional block attention module (CBAM) to the YOLOv5 backbone. The interference of underwater creature characteristics on object characteristics is decreased, and the output of the backbone network to object information is enhanced. In addition, the self-adaptive global histogram stretching algorithm (SAGHS) is designed to eliminate the degradation problems such as low contrast and color loss caused by underwater environmental information to better restore image quality. Extensive experiments and comprehensive evaluation on the URPC2021 benchmark dataset demonstrate the effectiveness and adaptivity of our methods. Beyond that, this paper conducts an exhaustive analysis of the role of training data on performance.
Recently, semantic search has been successfully applied to e-commerce product search and the learned semantic space(s) for query and product encoding are expected to generalize to unseen queries or products. Yet, whether generalization can conveniently emerge has not been thoroughly studied in the domain thus far. In this paper, we examine several general-domain and domain-specific pre-trained Roberta variants and discover that general-domain fine-tuning does not help generalization, which aligns with the discovery of prior art. Proper domain-specific fine-tuning with clickstream data can lead to better model generalization, based on a bucketed analysis of a publicly available manual annotated query-product pair da