Hao Zhou

Long Zhao, Nitesh B. Gundavarapu, Liangzhe Yuan et al. · deepmind

We introduce VideoPrism, a general-purpose video encoder that tackles diverse video understanding tasks with a single frozen model. We pretrain VideoPrism on a heterogeneous corpus containing 36M high-quality video-caption pairs and 582M video clips with noisy parallel text (e.g., ASR transcripts). The pretraining approach improves upon masked autoencoding by global-local distillation of semantic video embeddings and a token shuffling scheme, enabling VideoPrism to focus primarily on the video modality while leveraging the invaluable text associated with videos. We extensively test VideoPrism on four broad groups of video understanding tasks, from web video question answering to CV for science, achieving state-of-the-art performance on 31 out of 33 video understanding benchmarks. Our models are released at https://github.com/google-deepmind/videoprism.

Jun Yu, Hao Zhou, Yibing Zhan et al.

Unsupervised cross-modal hashing (UCMH) has become a hot topic recently. Current UCMH focuses on exploring data similarities. However, current UCMH methods calculate the similarity between two data, mainly relying on the two data's cross-modal features. These methods suffer from inaccurate similarity problems that result in a suboptimal retrieval Hamming space, because the cross-modal features between the data are not sufficient to describe the complex data relationships, such as situations where two data have different feature representations but share the inherent concepts. In this paper, we devise a deep graph-neighbor coherence preserving network (DGCPN). Specifically, DGCPN stems from graph models and explores graph-neighbor coherence by consolidating the information between data and their neighbors. DGCPN regulates comprehensive similarity preserving losses by exploiting three types of data similarities (i.e., the graph-neighbor coherence, the coexistent similarity, and the intra- and inter-modality consistency) and designs a half-real and half-binary optimization strategy to reduce the quantization errors during hashing. Essentially, DGCPN addresses the inaccurate similarity problem by exploring and exploiting the data's intrinsic relationships in a graph. We conduct extensive experiments on three public UCMH datasets. The experimental results demonstrate the superiority of DGCPN, e.g., by improving the mean average precision from 0.722 to 0.751 on MIRFlickr-25K using 64-bit hashing codes to retrieve texts from images. We will release the source code package and the trained model on https://github.com/Atmegal/DGCPN.

Kanzhi Cheng, Yantao Li, Fangzhi Xu et al.

Chain-of-thought (CoT) has proven to improve the reasoning capability of large language models (LLMs). However, due to the complexity of multimodal scenarios and the difficulty in collecting high-quality CoT data, CoT reasoning in multimodal LLMs has been largely overlooked. To this end, we propose a simple yet effective self-training framework, R3V, which iteratively enhances the model's Vision-language Reasoning by Reflecting on CoT Rationales. Our framework consists of two interleaved parts: (1) iteratively bootstrapping positive and negative solutions for reasoning datasets, and (2) reflection on rationale for learning from mistakes. Specifically, we introduce the self-refine and self-select losses, enabling the model to refine flawed rationale and derive the correct answer by comparing rationale candidates. Experiments on a wide range of vision-language tasks show that R3V consistently improves multimodal LLM reasoning, achieving a relative improvement of 23 to 60 percent over GPT-distilled baselines. Additionally, our approach supports self-reflection on generated solutions, further boosting performance through test-time computation.

Chen Chen, Guowu Yuan, Hao Zhou et al.

High-voltage transmission lines are located far from the road, resulting in inconvenient inspection work and rising maintenance costs. Intelligent inspection of power transmission lines has become increasingly important. However, subsequent intelligent inspection relies on accurately detecting various key components. Due to the low detection accuracy of key components in transmission line image inspection, this paper proposed an improved object detection model based on the YOLOv5s (You Only Look Once Version 5 Small) model to improve the detection accuracy of key components of transmission lines. According to the characteristics of the power grid inspection image, we first modify the distance measurement in the k-means clustering to improve the anchor matching of the YOLOv5s model. Then, we add the convolutional block attention module (CBAM) attention mechanism to the backbone network to improve accuracy. Finally, we apply the focal loss function to reduce the impact of class imbalance. Our improved method's mAP (mean average precision) reached 98.1%, the precision reached 97.5%, the recall reached 94.4%, and the detection rate reached 84.8 FPS (frames per second). The experimental results show that our improved model improves detection accuracy and has performance advantages over other models.

Kun Ouyang, Yuanxin Liu, Linli Yao et al.

Video reasoning, which requires multi-step deduction across frames, remains a major challenge for multimodal large language models (MLLMs). While reinforcement learning (RL)-based methods enhance reasoning capabilities, they often rely on text-only chains that yield ungrounded or hallucinated conclusions. Conversely, frame-retrieval approaches introduce visual grounding, yet still struggle with inaccurate evidence localization. To address these limitations, we present Conan, a framework for evidence-grounded multi-step video reasoning. Conan identifies context and evidence frames, reasons over cross-frame clues, and adaptively decides when to conclude or explore further. To achieve this, we 1) construct Conan-91K, a large-scale dataset of automatically generated reasoning traces that include frame identification, evidence reasoning, and action decision, and 2) design a multi-stage progressive cold-start strategy combined with an Identification-Reasoning-Action (AIR) RLVR training framework to progressively incentivize multi-step visual reasoning. Extensive experiments on six multi-step reasoning benchmarks demonstrate that Conan surpasses the baseline Qwen2.5-VL-7B-Instruct by an average of over 10% in accuracy, achieving state-of-the-art performance. Furthermore, Conan generalizes effectively to long video understanding tasks, validating its strong scalability and robustness.

Zhenyue Wang, Guowu Yuan, Hao Zhou et al.

Insulators are crucial insulation components and structural supports in power grids, playing a vital role in the transmission lines. Due to temperature fluctuations, internal stress, or damage from hail, insulators are prone to injury. Automatic detection of damaged insulators faces challenges such as diverse types, small defect targets, and complex backgrounds and shapes. Most research for detecting insulator defects has focused on a single defect type or a specific material. However, the insulators in the grid's transmission lines have different colors and materials. Various insulator defects coexist, and the existing methods have difficulty meeting the practical application requirements. Current methods suffer from low detection accuracy and mAP0.5 cannot meet application requirements. This paper proposes an improved YOLOv7 model for multi-type insulator defect detection. First, our model replaces the SPPCSPC module with the RFB module to enhance the network's feature extraction capability. Second, a CA mechanism is introduced into the head part to enhance the network's feature representation ability and to improve detection accuracy. Third, a WIoU loss function is employed to address the low-quality samples hindering model generalization during training, thereby improving the model's overall performance. The experimental results indicate that the proposed model exhibits enhancements across various performance metrics. Specifically, there is a 1.6% advancement in mAP_0.5, a corresponding 1.6% enhancement in mAP_0.5:0.95, a 1.3% elevation in precision, and a 1% increase in recall. Moreover, the model achieves parameter reduction by 3.2 million, leading to a decrease of 2.5 GFLOPS in computational cost. Notably, there is also an improvement of 2.81 milliseconds in single-image detection speed.

Lean Wang, Lei Li, Damai Dai et al.

In-context learning (ICL) emerges as a promising capability of large language models (LLMs) by providing them with demonstration examples to perform diverse tasks. However, the underlying mechanism of how LLMs learn from the provided context remains under-explored. In this paper, we investigate the working mechanism of ICL through an information flow lens. Our findings reveal that label words in the demonstration examples function as anchors: (1) semantic information aggregates into label word representations during the shallow computation layers' processing; (2) the consolidated information in label words serves as a reference for LLMs' final predictions. Based on these insights, we introduce an anchor re-weighting method to improve ICL performance, a demonstration compression technique to expedite inference, and an analysis framework for diagnosing ICL errors in GPT2-XL. The promising applications of our findings again validate the uncovered ICL working mechanism and pave the way for future studies.

Koutilya PNVR, Hao Zhou, David Jacobs

We propose a novel method for combining synthetic and real images when training networks to determine geometric information from a single image. We suggest a method for mapping both image types into a single, shared domain. This is connected to a primary network for end-to-end training. Ideally, this results in images from two domains that present shared information to the primary network. Our experiments demonstrate significant improvements over the state-of-the-art in two important domains, surface normal estimation of human faces and monocular depth estimation for outdoor scenes, both in an unsupervised setting.

Hao Zhou, Ke Chen

Speech emotion recognition plays an important role in building more intelligent and human-like agents. Due to the difficulty of collecting speech emotional data, an increasingly popular solution is leveraging a related and rich source corpus to help address the target corpus. However, domain shift between the corpora poses a serious challenge, making domain shift adaptation difficult to function even on the recognition of positive/negative emotions. In this work, we propose class-wise adversarial domain adaptation to address this challenge by reducing the shift for all classes between different corpora. Experiments on the well-known corpora EMODB and Aibo demonstrate that our method is effective even when only a very limited number of target labeled examples are provided.

Hui Ding, Hao Zhou, Shaohua Kevin Zhou et al.

Humans focus attention on different face regions when recognizing face attributes. Most existing face attribute classification methods use the whole image as input. Moreover, some of these methods rely on fiducial landmarks to provide defined face parts. In this paper, we propose a cascade network that simultaneously learns to localize face regions specific to attributes and performs attribute classification without alignment. First, a weakly-supervised face region localization network is designed to automatically detect regions (or parts) specific to attributes. Then multiple part-based networks and a whole-image-based network are separately constructed and combined together by the region switch layer and attribute relation layer for final attribute classification. A multi-net learning method and hint-based model compression is further proposed to get an effective localization model and a compact classification model, respectively. Our approach achieves significantly better performance than state-of-the-art methods on unaligned CelebA dataset, reducing the classification error by 30.9%.

Hao Zhou, Jin Sun, Yaser Yacoob et al.

Lighting estimation from face images is an important task and has applications in many areas such as image editing, intrinsic image decomposition, and image forgery detection. We propose to train a deep Convolutional Neural Network (CNN) to regress lighting parameters from a single face image. Lacking massive ground truth lighting labels for face images in the wild, we use an existing method to estimate lighting parameters, which are treated as ground truth with unknown noises. To alleviate the effect of such noises, we utilize the idea of Generative Adversarial Networks (GAN) and propose a Label Denoising Adversarial Network (LDAN) to make use of synthetic data with accurate ground truth to help train a deep CNN for lighting regression on real face images. Experiments show that our network outperforms existing methods in producing consistent lighting parameters of different faces under similar lighting conditions. Moreover, our method is 100,000 times faster in execution time than prior optimization-based lighting estimation approaches.

Soumyadip Sengupta, Hao Zhou, Walter Forkel et al.

We introduce a new, integrated approach to uncalibrated photometric stereo. We perform 3D reconstruction of Lambertian objects using multiple images produced by unknown, directional light sources. We show how to formulate a single optimization that includes rank and integrability constraints, allowing also for missing data. We then solve this optimization using the Alternate Direction Method of Multipliers (ADMM). We conduct extensive experimental evaluation on real and synthetic data sets. Our integrated approach is particularly valuable when performing photometric stereo using as few as 4-6 images, since the integrability constraint is capable of improving estimation of the linear subspace of possible solutions. We show good improvements over prior work in these cases.