Zhenyu Wang

Zhenyu Wang, Yali Li, Xi Chen et al. · deepmind

In this paper, we formally address universal object detection, which aims to detect every scene and predict every category. The dependence on human annotations, the limited visual information, and the novel categories in the open world severely restrict the universality of traditional detectors. We propose UniDetector, a universal object detector that has the ability to recognize enormous categories in the open world. The critical points for the universality of UniDetector are: 1) it leverages images of multiple sources and heterogeneous label spaces for training through the alignment of image and text spaces, which guarantees sufficient information for universal representations. 2) it generalizes to the open world easily while keeping the balance between seen and unseen classes, thanks to abundant information from both vision and language modalities. 3) it further promotes the generalization ability to novel categories through our proposed decoupling training manner and probability calibration. These contributions allow UniDetector to detect over 7k categories, the largest measurable category size so far, with only about 500 classes participating in training. Our UniDetector behaves the strong zero-shot generalization ability on large-vocabulary datasets like LVIS, ImageNetBoxes, and VisualGenome - it surpasses the traditional supervised baselines by more than 4\% on average without seeing any corresponding images. On 13 public detection datasets with various scenes, UniDetector also achieves state-of-the-art performance with only a 3\% amount of training data.

You Shen, Yunzhou Zhang, Yanmin Wu et al. · pku

The progress of LiDAR-based 3D object detection has significantly enhanced developments in autonomous driving and robotics. However, due to the limitations of LiDAR sensors, object shapes suffer from deterioration in occluded and distant areas, which creates a fundamental challenge to 3D perception. Existing methods estimate specific 3D shapes and achieve remarkable performance. However, these methods rely on extensive computation and memory, causing imbalances between accuracy and real-time performance. To tackle this challenge, we propose a novel LiDAR-based 3D object detection model named BSH-Det3D, which applies an effective way to enhance spatial features by estimating complete shapes from a bird's eye view (BEV). Specifically, we design the Pillar-based Shape Completion (PSC) module to predict the probability of occupancy whether a pillar contains object shapes. The PSC module generates a BEV shape heatmap for each scene. After integrating with heatmaps, BSH-Det3D can provide additional information in shape deterioration areas and generate high-quality 3D proposals. We also design an attention-based densification fusion module (ADF) to adaptively associate the sparse features with heatmaps and raw points. The ADF module integrates the advantages of points and shapes knowledge with negligible overheads. Extensive experiments on the KITTI benchmark achieve state-of-the-art (SOTA) performance in terms of accuracy and speed, demonstrating the efficiency and flexibility of BSH-Det3D. The source code is available on https://github.com/mystorm16/BSH-Det3D.

Zhenyu Wang, Li Wan, Biqiao Zhang et al. · amazon-science, meta-ai

A keyword spotting (KWS) engine that is continuously running on device is exposed to various speech signals that are usually unseen before. It is a challenging problem to build a small-footprint and high-performing KWS model with robustness under different acoustic environments. In this paper, we explore how to effectively apply adversarial examples to improve KWS robustness. We propose datasource-aware disentangled learning with adversarial examples to reduce the mismatch between the original and adversarial data as well as the mismatch across original training datasources. The KWS model architecture is based on depth-wise separable convolution and a simple attention module. Experimental results demonstrate that the proposed learning strategy improves false reject rate by $40.31%$ at $1%$ false accept rate on the internal dataset, compared to the strongest baseline without using adversarial examples. Our best-performing system achieves $98.06%$ accuracy on the Google Speech Commands V1 dataset.

Zhenyu Wang, Yali Li, Shengjin Wang

Current instance segmentation methods rely heavily on pixel-level annotated images. The huge cost to obtain such fully-annotated images restricts the dataset scale and limits the performance. In this paper, we formally address semi-supervised instance segmentation, where unlabeled images are employed to boost the performance. We construct a framework for semi-supervised instance segmentation by assigning pixel-level pseudo labels. Under this framework, we point out that noisy boundaries associated with pseudo labels are double-edged. We propose to exploit and resist them in a unified manner simultaneously: 1) To combat the negative effects of noisy boundaries, we propose a noise-tolerant mask head by leveraging low-resolution features. 2) To enhance the positive impacts, we introduce a boundary-preserving map for learning detailed information within boundary-relevant regions. We evaluate our approach by extensive experiments. It behaves extraordinarily, outperforming the supervised baseline by a large margin, more than 6% on Cityscapes, 7% on COCO and 4.5% on BDD100k. On Cityscapes, our method achieves comparable performance by utilizing only 30% labeled images.

Zhenyu Wang, John H. L. Hansen

Adapting speaker recognition systems to new environments is a widely-used technique to improve a well-performing model learned from large-scale data towards a task-specific small-scale data scenarios. However, previous studies focus on single domain adaptation, which neglects a more practical scenario where training data are collected from multiple acoustic domains needed in forensic scenarios. Audio analysis for forensic speaker recognition offers unique challenges in model training with multi-domain training data due to location/scenario uncertainty and diversity mismatch between reference and naturalistic field recordings. It is also difficult to directly employ small-scale domain-specific data to train complex neural network architectures due to domain mismatch and performance loss. Fine-tuning is a commonly-used method for adaptation in order to retrain the model with weights initialized from a well-trained model. Alternatively, in this study, three novel adaptation methods based on domain adversarial training, discrepancy minimization, and moment-matching approaches are proposed to further promote adaptation performance across multiple acoustic domains. A comprehensive set of experiments are conducted to demonstrate that: 1) diverse acoustic environments do impact speaker recognition performance, which could advance research in audio forensics, 2) domain adversarial training learns the discriminative features which are also invariant to shifts between domains, 3) discrepancy-minimizing adaptation achieves effective performance simultaneously across multiple acoustic domains, and 4) moment-matching adaptation along with dynamic distribution alignment also significantly promotes speaker recognition performance on each domain, especially for the LENA-field domain with noise compared to all other systems.

Zhenyu Wang, Shuyu Kong, Li Wan et al.

Existing keyword spotting (KWS) systems primarily rely on predefined keyword phrases. However, the ability to recognize customized keywords is crucial for tailoring interactions with intelligent devices. In this paper, we present a novel Query-by-Example (QbyE) KWS system that employs spectral-temporal graph attentive pooling and multi-task learning. This framework aims to effectively learn speaker-invariant and linguistic-informative embeddings for QbyE KWS tasks. Within this framework, we investigate three distinct network architectures for encoder modeling: LiCoNet, Conformer and ECAPA_TDNN. The experimental results on a substantial internal dataset of $629$ speakers have demonstrated the effectiveness of the proposed QbyE framework in maximizing the potential of simpler models such as LiCoNet. Particularly, LiCoNet, which is 13x more efficient, achieves comparable performance to the computationally intensive Conformer model (1.98% vs. 1.63\% FRR at 0.3 FAs/Hr).

Zhenyu Wang, John H. L. Hansen

Audio analysis for forensic speaker verification offers unique challenges in system performance due in part to data collected in naturalistic field acoustic environments where location/scenario uncertainty is common in the forensic data collection process. Forensic speech data as potential evidence can be obtained in random naturalistic environments resulting in variable data quality. Speech samples may include variability due to vocal efforts such as yelling over 911 emergency calls, whereas others might be whisper or situational stressed voice in a field location or interview room. Such speech variability consists of intrinsic and extrinsic characteristics and makes forensic speaker verification a complicated and daunting task. Extrinsic properties include recording equipment such as microphone type and placement, ambient noise, room configuration including reverberation, and other environmental scenario-based issues. Some factors, such as noise and non-target speech, will impact the verification system performance by their mere presence. To investigate the impact of field acoustic environments, we performed a speaker verification study based on the CRSS-Forensic corpus with audio collected from 8 field locations including police interviews. This investigation includes an analysis of the impact of seven unseen acoustic environments on speaker verification system performance using an x-Vector system.

Rui Peng, Rongjie Wang, Zhenyu Wang et al.

Depth estimation is solved as a regression or classification problem in existing learning-based multi-view stereo methods. Although these two representations have recently demonstrated their excellent performance, they still have apparent shortcomings, e.g., regression methods tend to overfit due to the indirect learning cost volume, and classification methods cannot directly infer the exact depth due to its discrete prediction. In this paper, we propose a novel representation, termed Unification, to unify the advantages of regression and classification. It can directly constrain the cost volume like classification methods, but also realize the sub-pixel depth prediction like regression methods. To excavate the potential of unification, we design a new loss function named Unified Focal Loss, which is more uniform and reasonable to combat the challenge of sample imbalance. Combining these two unburdened modules, we present a coarse-to-fine framework, that we call UniMVSNet. The results of ranking first on both DTU and Tanks and Temples benchmarks verify that our model not only performs the best but also has the best generalization ability.

Zhenyu Wang, Yali Li, Ye Guo et al.

In this paper, we delve into semi-supervised object detection where unlabeled images are leveraged to break through the upper bound of fully-supervised object detection models. Previous semi-supervised methods based on pseudo labels are severely degenerated by noise and prone to overfit to noisy labels, thus are deficient in learning different unlabeled knowledge well. To address this issue, we propose a data-uncertainty guided multi-phase learning method for semi-supervised object detection. We comprehensively consider divergent types of unlabeled images according to their difficulty levels, utilize them in different phases and ensemble models from different phases together to generate ultimate results. Image uncertainty guided easy data selection and region uncertainty guided RoI Re-weighting are involved in multi-phase learning and enable the detector to concentrate on more certain knowledge. Through extensive experiments on PASCAL VOC and MS COCO, we demonstrate that our method behaves extraordinarily compared to baseline approaches and outperforms them by a large margin, more than 3% on VOC and 2% on COCO.

Zhenyu Wang, Wei Xia, John H. L. Hansen

Forensic audio analysis for speaker verification offers unique challenges due to location/scenario uncertainty and diversity mismatch between reference and naturalistic field recordings. The lack of real naturalistic forensic audio corpora with ground-truth speaker identity represents a major challenge in this field. It is also difficult to directly employ small-scale domain-specific data to train complex neural network architectures due to domain mismatch and loss in performance. Alternatively, cross-domain speaker verification for multiple acoustic environments is a challenging task which could advance research in audio forensics. In this study, we introduce a CRSS-Forensics audio dataset collected in multiple acoustic environments. We pre-train a CNN-based network using the VoxCeleb data, followed by an approach which fine-tunes part of the high-level network layers with clean speech from CRSS-Forensics. Based on this fine-tuned model, we align domain-specific distributions in the embedding space with the discrepancy loss and maximum mean discrepancy (MMD). This maintains effective performance on the clean set, while simultaneously generalizes the model to other acoustic domains. From the results, we demonstrate that diverse acoustic environments affect the speaker verification performance, and that our proposed approach of cross-domain adaptation can significantly improve the results in this scenario.

Fei Tang, Wanling Gao, Jianfeng Zhan et al.

Earlier-stage evaluations of a new AI architecture/system need affordable benchmarks. Only using a few AI component benchmarks like MLPerfalone in the other stages may lead to misleading conclusions. Moreover, the learning dynamics are not well understood, and the benchmarks' shelf-life is short. This paper proposes a balanced benchmarking methodology. We use real-world benchmarks to cover the factors space that impacts the learning dynamics to the most considerable extent. After performing an exhaustive survey on Internet service AI domains, we identify and implement nineteen representative AI tasks with state-of-the-art models. For repeatable performance ranking (RPR subset) and workload characterization (WC subset), we keep two subsets to a minimum for affordability. We contribute by far the most comprehensive AI training benchmark suite. The evaluations show: (1) AIBench Training (v1.1) outperforms MLPerfTraining (v0.7) in terms of diversity and representativeness of model complexity, computational cost, convergent rate, computation, and memory access patterns, and hotspot functions; (2) Against the AIBench full benchmarks, its RPR subset shortens the benchmarking cost by 64%, while maintaining the primary workload characteristics; (3) The performance ranking shows the single-purpose AI accelerator like TPU with the optimized TensorFlowframework performs better than that of GPUs while losing the latter's general support for various AI models. The specification, source code, and performance numbers are available from the AIBench homepage https://www.benchcouncil.org/aibench-training/index.html.

Hao Li, Yongguo Kang, Zhenyu Wang

We present EMPHASIS, an emotional phoneme-based acoustic model for speech synthesis system. EMPHASIS includes a phoneme duration prediction model and an acoustic parameter prediction model. It uses a CBHG-based regression network to model the dependencies between linguistic features and acoustic features. We modify the input and output layer structures of the network to improve the performance. For the linguistic features, we apply a feature grouping strategy to enhance emotional and prosodic features. The acoustic parameters are designed to be suitable for the regression task and waveform reconstruction. EMPHASIS can synthesize speech in real-time and generate expressive interrogative and exclamatory speech with high audio quality. EMPHASIS is designed to be a multi-lingual model and can synthesize Mandarin-English speech for now. In the experiment of emotional speech synthesis, it achieves better subjective results than other real-time speech synthesis systems.