Automatic colorization of anime line drawing has attracted much attention in recent years since it can substantially benefit the animation industry. User-hint based methods are the mainstream approach for line drawing colorization, while reference-based methods offer a more intuitive approach. Nevertheless, although reference-based methods can improve feature aggregation of the reference image and the line drawing, the colorization results are not compelling in terms of color consistency or semantic correspondence. In this paper, we introduce an attention-based model for anime line drawing colorization, in which a channel-wise and spatial-wise Convolutional Attention module is used to improve the ability of the encoder for feature extraction and key area perception, and a Stop-Gradient Attention module with cross-attention and self-attention is used to tackle the cross-domain long-range dependency problem. Extensive experiments show that our method outperforms other SOTA methods, with more accurate line structure and semantic color information.

This work explores an efficient approach to establish a foundational video-text model for tasks including open-vocabulary video classification, text-to-video retrieval, video captioning and video question-answering. We present VideoCoCa that reuses a pretrained image-text contrastive captioner (CoCa) model and adapt it to video-text tasks with minimal extra training. While previous works adapt image-text models with various cross-frame fusion modules (for example, cross-frame attention layer or perceiver resampler) and finetune the modified architecture on video-text data, we surprisingly find that the generative attentional pooling and contrastive attentional pooling layers in the image-text CoCa design are instantly adaptable to ``flattened frame embeddings'', yielding a strong zero-shot transfer baseline for many video-text tasks. Specifically, the frozen image encoder of a pretrained image-text CoCa takes each video frame as inputs and generates \(N\) token embeddings per frame for totally \(T\) video frames. We flatten \(N \times T\) token embeddings as a long sequence of frozen video representation and apply CoCa's generative attentional pooling and contrastive attentional pooling on top. All model weights including pooling layers are directly loaded from an image-text CoCa pretrained model. Without any video or video-text data, VideoCoCa's zero-shot transfer baseline already achieves state-of-the-art results on zero-shot video classification on Kinetics 400/600/700, UCF101, HMDB51, and Charades, as well as zero-shot text-to-video retrieval on MSR-VTT and ActivityNet Captions. We also explore lightweight finetuning on top of VideoCoCa, and achieve strong results on video question-answering (iVQA, MSRVTT-QA, MSVD-QA) and video captioning (MSR-VTT, ActivityNet, Youcook2). Our approach establishes a simple and effective video-text baseline for future research.

Large deep learning models have achieved remarkable success in many scenarios. However, training large models is usually challenging, e.g., due to the high computational cost, the unstable and painfully slow optimization procedure, and the vulnerability to overfitting. To alleviate these problems, this work studies a divide-and-conquer strategy, i.e., dividing a large model into smaller modules, training them independently, and reassembling the trained modules to obtain the target model. This approach is promising since it avoids directly training large models from scratch. Nevertheless, implementing this idea is non-trivial, as it is difficult to ensure the compatibility of the independently trained modules. In this paper, we present an elegant solution to address this issue, i.e., we introduce a global, shared meta model to implicitly link all the modules together. This enables us to train highly compatible modules that collaborate effectively when they are assembled together. We further propose a module incubation mechanism that enables the meta model to be designed as an extremely shallow network. As a result, the additional overhead introduced by the meta model is minimalized. Though conceptually simple, our method significantly outperforms end-to-end (E2E) training in terms of both final accuracy and training efficiency. For example, on top of ViT-Huge, it improves the accuracy by 2.7% compared to the E2E baseline on ImageNet-1K, while saving the training cost by 43% in the meantime. Code is available at https://github.com/LeapLabTHU/Model-Assembling.

The security of artificial intelligence (AI) is an important research area towards safe, reliable, and trustworthy AI systems. To accelerate the research on AI security, the Artificial Intelligence Security Competition (AISC) was organized by the Zhongguancun Laboratory, China Industrial Control Systems Cyber Emergency Response Team, Institute for Artificial Intelligence, Tsinghua University, and RealAI as part of the Zhongguancun International Frontier Technology Innovation Competition (https://www.zgc-aisc.com/en). The competition consists of three tracks, including Deepfake Security Competition, Autonomous Driving Security Competition, and Face Recognition Security Competition. This report will introduce the competition rules of these three tracks and the solutions of top-ranking teams in each track.

Vision-language foundation models pretrained on large-scale data provide a powerful tool for many visual understanding tasks. Notably, many vision-language models build two encoders (visual and textual) that can map two modalities into the same embedding space. As a result, the learned representations achieve good zero-shot performance on tasks like image classification. However, when there are only a few examples per category, the potential of large vision-language models is often underperformed, mainly due to the gap between a large number of parameters and a relatively small amount of training data. This paper shows that we can significantly improve the performance of few-shot classification by using the category names to initialize the classification head. More interestingly, we can borrow the non-perfect category names, or even names from a foreign language, to improve the few-shot classification performance compared with random initialization. With the proposed category name initialization method, our model obtains the state-of-the-art performance on a number of few-shot image classification benchmarks (e.g., 87.37\% on ImageNet and 96.08\% on Stanford Cars, both using five-shot learning). We also investigate and analyze when the benefit of category names diminishes and how to use distillation to improve the performance of smaller models, providing guidance for future research.

Pretrained language models (PLMs) have motivated research on what kinds of knowledge these models learn. Fill-in-the-blanks problem (e.g., cloze tests) is a natural approach for gauging such knowledge. BioLAMA generates prompts for biomedical factual knowledge triples and uses the Top-k accuracy metric to evaluate different PLMs' knowledge. However, existing research has shown that such prompt-based knowledge probing methods can only probe a lower bound of knowledge. Many factors like prompt-based probing biases make the LAMA benchmark unreliable and unstable. This problem is more prominent in BioLAMA. The severe long-tailed distribution in vocabulary and large-N-M relation make the performance gap between LAMA and BioLAMA remain notable. To address these, we introduce context variance into the prompt generation and propose a new rank-change-based evaluation metric. Different from the previous known-unknown evaluation criteria, we propose the concept of "Misunderstand" in LAMA for the first time. Through experiments on 12 PLMs, our context variance prompts and Understand-Confuse-Misunderstand (UCM) metric makes BioLAMA more friendly to large-N-M relations and rare relations. We also conducted a set of control experiments to disentangle "understand" from just "read and copy".

Various depth estimation models are now widely used on many mobile and IoT devices for image segmentation, bokeh effect rendering, object tracking and many other mobile tasks. Thus, it is very crucial to have efficient and accurate depth estimation models that can run fast on low-power mobile chipsets. In this Mobile AI challenge, the target was to develop deep learning-based single image depth estimation solutions that can show a real-time performance on IoT platforms and smartphones. For this, the participants used a large-scale RGB-to-depth dataset that was collected with the ZED stereo camera capable to generated depth maps for objects located at up to 50 meters. The runtime of all models was evaluated on the Raspberry Pi 4 platform, where the developed solutions were able to generate VGA resolution depth maps at up to 27 FPS while achieving high fidelity results. All models developed in the challenge are also compatible with any Android or Linux-based mobile devices, their detailed description is provided in this paper.

In natural language processing (NLP), the context of a word or sentence plays an essential role. Contextual information such as the semantic representation of a passage or historical dialogue forms an essential part of a conversation and a precise understanding of the present phrase or sentence. However, the standard attention mechanisms typically generate weights using query and key but ignore context, forming a Bi-Attention framework, despite their great success in modeling sequence alignment. This Bi-Attention mechanism does not explicitly model the interactions between the contexts, queries and keys of target sequences, missing important contextual information and resulting in poor attention performance. Accordingly, a novel and general triple-attention (Tri-Attention) framework expands the standard Bi-Attention mechanism and explicitly interacts query, key, and context by incorporating context as the third dimension in calculating relevance scores. Four variants of Tri-Attention are generated by expanding the two-dimensional vector-based additive, dot-product, scaled dot-product, and bilinear operations in Bi-Attention to the tensor operations for Tri-Attention. Extensive experiments on three NLP tasks demonstrate that Tri-Attention outperforms about 30 state-of-the-art non-attention, standard Bi-Attention, contextual Bi-Attention approaches and pretrained neural language models1.

Generative Adversarial Networks (GANs) typically suffer from overfitting when limited training data is available. To facilitate GAN training, current methods propose to use data-specific augmentation techniques. Despite the effectiveness, it is difficult for these methods to scale to practical applications. In this work, we present ScoreMix, a novel and scalable data augmentation approach for various image synthesis tasks. We first produce augmented samples using the convex combinations of the real samples. Then, we optimize the augmented samples by minimizing the norms of the data scores, i.e., the gradients of the log-density functions. This procedure enforces the augmented samples close to the data manifold. To estimate the scores, we train a deep estimation network with multi-scale score matching. For different image synthesis tasks, we train the score estimation network using different data. We do not require the tuning of the hyperparameters or modifications to the network architecture. The ScoreMix method effectively increases the diversity of data and reduces the overfitting problem. Moreover, it can be easily incorporated into existing GAN models with minor modifications. Experimental results on numerous tasks demonstrate that GAN models equipped with the ScoreMix method achieve significant improvements.

We present a novel neural surface reconstruction method called NeuralRoom for reconstructing room-sized indoor scenes directly from a set of 2D images. Recently, implicit neural representations have become a promising way to reconstruct surfaces from multiview images due to their high-quality results and simplicity. However, implicit neural representations usually cannot reconstruct indoor scenes well because they suffer severe shape-radiance ambiguity. We assume that the indoor scene consists of texture-rich and flat texture-less regions. In texture-rich regions, the multiview stereo can obtain accurate results. In the flat area, normal estimation networks usually obtain a good normal estimation. Based on the above observations, we reduce the possible spatial variation range of implicit neural surfaces by reliable geometric priors to alleviate shape-radiance ambiguity. Specifically, we use multiview stereo results to limit the NeuralRoom optimization space and then use reliable geometric priors to guide NeuralRoom training. Then the NeuralRoom would produce a neural scene representation that can render an image consistent with the input training images. In addition, we propose a smoothing method called perturbation-residual restrictions to improve the accuracy and completeness of the flat region, which assumes that the sampling points in a local surface should have the same normal and similar distance to the observation center. Experiments on the ScanNet dataset show that our method can reconstruct the texture-less area of indoor scenes while maintaining the accuracy of detail. We also apply NeuralRoom to more advanced multiview reconstruction algorithms and significantly improve their reconstruction quality.