AAAI 2024 Papers

The superior performances of pre-trained foundation models in various visual tasks underscore their potential to enhance the 2D models' open-vocabulary ability. Existing methods explore analogous applications in the 3D space. However, most of them only center around knowledge extraction from singular foundation models, which limits the open-vocabulary ability of 3D models. We hypothesize that leveraging complementary pre-trained knowledge from various foundation models can improve knowledge transfer from 2D pre-trained visual language models to the 3D space. In this work, we propose FM-OV3D, a method of Foundation Model-based Cross-modal Knowledge Blending for Open-Vocabulary 3D Detection, which improves the open-vocabulary localization and recognition abilities of 3D model by blending knowledge from multiple pre-trained foundation models, achieving true open-vocabulary without facing constraints from original 3D datasets. Specifically, to learn the open-vocabulary 3D localization ability, we adopt the open-vocabulary localization knowledge of the Grounded-Segment-Anything model. For open-vocabulary 3D recognition ability, We leverage the knowledge of generative foundation models, including GPT-3 and Stable Diffusion models, and cross-modal discriminative models like CLIP. The experimental results on two popular benchmarks for open-vocabulary 3D object detection show that our model efficiently learns knowledge from multiple foundation models to enhance the open-vocabulary ability of the 3D model and successfully achieves state-of-the-art performance in open-vocabulary 3D object detection tasks. Code is released at https://github.com/dmzhang0425/FM-OV3D.git.

Spreadsheets are a vital tool for end-user data management. Using large language models for formula authoring assistance in these environments can be difficult, as these models are expensive to train and challenging to deploy due to their size (up to billions of parameters). We present FLAME, a transformer-based model trained exclusively on Excel formulas that leverages domain insights to achieve competitive performance while being substantially smaller (60M parameters) and training on two orders of magnitude less data. We curate a training dataset using sketch deduplication, introduce an Excel-specific formula tokenizer, and use domain-specific versions of masked span prediction and noisy auto-encoding as pre-training objectives. We evaluate FLAME on formula repair, formula completion, and similarity-based formula retrieval. FLAME can outperform much larger models, such as the Davinci (175B) and Cushman (12B) variants of Codex and CodeT5 (220M), in 10 of 14 evaluation settings for the repair and completion tasks. For formula retrieval, FLAME outperforms CodeT5, CodeBERT, and GraphCodeBERT.

Although deep generative models have greatly improved one-shot video-driven talking head generation, few studies address fine-grained controllable facial expression editing, which is crucial for practical applications. Existing methods rely on a fixed set of predefined discrete emotion labels or simply copy expressions from input videos. This is limiting as expressions are complex, and methods using only emotion labels cannot generate fine-grained, accurate or mixed expressions. Generating talking head video with precise expressions is also difficult using 3D model-based approaches, as 3DMM only models facial movements and tends to produce deviations. In this paper, we propose a novel framework enabling fine-grained facial expression editing in talking face generation. Our goal is to achieve expression control by manipulating the intensities of individual facial Action Units (AUs) or groups. First, compared with existing methods which decouple the face into pose and expression, we propose a disentanglement scheme to isolates three components from the human face, namely, appearance, pose, and expression. Second, we propose to use input AUs to control muscle group intensities in the generated face, and integrate the AUs features with the disentangled expression latent code. Finally, we present a self-supervised training strategy with well-designed constraints. Experiments show our method achieves fine-grained expression control, produces high-quality talking head videos and outperforms baseline methods.

Multi-head-self-attention (MHSA)-equipped models have achieved notable performance in computer vision. Their computational complexity is proportional to quadratic numbers of pixels in input feature maps, resulting in slow processing, especially when dealing with high-resolution images. New types of token-mixer are proposed as an alternative to MHSA to circumvent this problem: an FFT-based token-mixer involves global operations similar to MHSA but with lower computational complexity. However, despite its attractive properties, the FFT-based token-mixer has not been carefully examined in terms of its compatibility with the rapidly evolving MetaFormer architecture. Here, we propose a novel token-mixer called Dynamic Filter and novel image recognition models, DFFormer and CDFFormer, to close the gaps above. The results of image classification and downstream tasks, analysis, and visualization show that our models are helpful. Notably, their throughput and memory efficiency when dealing with high-resolution image recognition is remarkable. Our results indicate that Dynamic Filter is one of the token-mixer options that should be seriously considered. The code is available at https://github.com/okojoalg/dfformer

Monocular 3D object detection usually adopts direct or hierarchical label supervision. Recently, the distillation supervision transfers the spatial knowledge from LiDAR- or stereo-based teacher networks to monocular detectors, but remaining the domain gap. To mitigate this issue and pursue adequate label manipulation, we exploit Foreground Depth map for feature-supervised monocular 3D object detection named FD3D, which develops the high-quality instructive intermediate features to conduct desirable auxiliary feature supervision with only the original image and annotation foreground object-wise depth map (AFOD) as input. Furthermore, we build up our instructive feature generation network to create instructive spatial features based on the sufficient correlation between image features and pre-processed AFOD, where AFOD provides the attention focus only on foreground objects to achieve clearer guidance in the detection task. Moreover, we apply the auxiliary feature supervision from the pixel and distribution level to achieve comprehensive spatial knowledge guidance. Extensive experiments demonstrate that our method achieves state-of-the-art performance on both the KITTI and nuScenes datasets, with no external data and no extra inference computational cost. We also conduct quantitative and qualitative studies to reveal the effectiveness of our designs.

Rearrangement operations form the crux of interactions between humans and their environment. The ability to generate natural, fluid sequences of this operation is of essential value in AR/VR and CG. Bridging a gap in the field, our study introduces FAVOR: a novel dataset for Full-body AR-driven Virtual Object Rearrangement that uniquely employs motion capture systems and AR eyeglasses. Comprising 3k diverse motion rearrangement sequences and 7.17 million interaction data frames, this dataset breaks new ground in research data. We also present a pipeline FAVORITE for producing digital human rearrangement motion sequences guided by instructions. Experimental results, both qualitative and quantitative, suggest that this dataset and pipeline deliver high-quality motion sequences. Our dataset, code, and appendix are available at https://kailinli.github.io/FAVOR.

Facial Expression Recognition (FER) is an extensively explored research problem in the domain of computer vision and artificial intelligence. FER, a supervised learning problem, requires significant training data representative of multiple socio-cultural demographic attributes. However, most of the FER dataset consists of images annotated by humans, which propagates individual and demographic biases. This work attempts to mitigate this bias using representation learning based on latent spaces, thereby increasing a deep learning model's fairness and overall accuracy.

Deep neural networks are known to be vulnerable to security risks due to the inherent transferable nature of adversarial examples. Despite the success of recent generative model-based attacks demonstrating strong transferability, it still remains a challenge to design an efficient attack strategy in a real-world strict black-box setting, where both the target domain and model architectures are unknown. In this paper, we seek to explore a feature contrastive approach in the frequency domain to generate adversarial examples that are robust in both cross-domain and cross-model settings. With that goal in mind, we propose two modules that are only employed during the training phase: a Frequency-Aware Domain Randomization (FADR) module to randomize domain-variant low- and high-range frequency components and a Frequency-Augmented Contrastive Learning (FACL) module to effectively separate domain-invariant mid-frequency features of clean and perturbed image. We demonstrate strong transferability of our generated adversarial perturbations through extensive cross-domain and cross-model experiments, while keeping the inference time complexity.

In the digital age, understanding the dynamics of information spread and opinion formation within networks is paramount. This research introduces an innovative framework that combines the principles of opinion dynamics with the strategic capabilities of Automated Planning. We have developed, to the best of our knowledge, the first-ever numeric PDDL tailored for opinion dynamics. Our tool empowers users to visualize intricate networks, simulate the evolution of opinions, and strategically influence that evolution to achieve specific outcomes. By harnessing Automated Planning techniques, our framework offers a nuanced approach to devise sequences of actions tailored to transition a network from its current opinion landscape to a desired state. This holistic approach provides insights into the intricate interplay of individual nodes within a network and paves the way for targeted interventions. Furthermore, the tool facilitates human-AI collaboration, enabling users to not only understand information spread but also devise practical strategies to mitigate potential harmful outcomes arising from it. Demo Video link - https://tinyurl.com/3k7bp99h

Information sharing through communication is essential for tackling complex multi-agent reinforcement learning tasks. Many existing multi-agent communication protocols can be viewed as instances of message passing graph neural networks (GNNs). However, due to the significantly limited expressive ability of the standard GNN method, the agent feature representations remain similar and indistinguishable even though the agents have different neighborhood structures. This further results in the homogenization of agent behaviors and reduces the capability to solve tasks effectively. In this paper, we propose a multi-agent communication protocol via identity-aware learning (IDEAL), which explicitly enhances the distinguishability of agent feature representations to break the diversity bottleneck. Specifically, IDEAL extends existing multi-agent communication protocols by inductively considering the agents' identities during the message passing process. To obtain expressive feature representations for a given agent, IDEAL first extracts the ego network centered around that agent and then performs multiple rounds of heterogeneous message passing, where different parameter sets are applied to the central agent and the other surrounding agents within the ego network. IDEAL fosters expressive communication between agents and generates distinguishable feature representations, which promotes action diversity and individuality emergence. Experimental results on various benchmarks demonstrate IDEAL can be flexibly integrated into various multi-agent communication methods and enhances the corresponding performance.

Human motion forecasting, with the goal of estimating future human behavior over a period of time, is a fundamental task in many real-world applications. However, existing works typically concentrate on foretelling the major joints of the human body without considering the delicate movements of the human hands. In practical applications, hand gesture plays an important role in human communication with the real world, and expresses the primary intention of human beings. In this work, we are the first to formulate whole-body human pose forecasting task, which jointly predicts future both body and gesture activities. Correspondingly, we propose a novel Encoding-Alignment-Interaction (EAI) framework that aims to predict both coarse (body joints) and fine-grained (gestures) activities collaboratively, enabling expressive and cross-facilitated forecasting of 3D whole-body human motions. Specifically, our model involves two key constituents: cross-context alignment (XCA) and cross-context interaction (XCI). Considering the heterogeneous information within the whole-body, XCA aims to align the latent features of various human components, while XCI focuses on effectively capturing the context interaction among the human components. We conduct extensive experiments on a newly-introduced large-scale benchmark and achieve state-of-the-art performance. The code is public for research purposes at https://github.com/Dingpx/EAI.

Deepfake technology has given rise to a spectrum of novel and compelling applications. Unfortunately, the widespread proliferation of high-fidelity fake videos has led to pervasive confusion and deception, shattering our faith that seeing is believing. One aspect that has been overlooked so far is that current deepfake detection approaches may easily fall into the trap of overfitting, focusing only on forgery clues within one or a few local regions. Moreover, existing works heavily rely on neural networks to extract forgery features, lacking theoretical constraints guaranteeing that sufficient forgery clues are extracted and superfluous features are eliminated. These deficiencies culminate in unsatisfactory accuracy and limited generalizability in real-life scenarios. In this paper, we try to tackle these challenges through three designs: (1) We present a novel framework to capture broader forgery clues by extracting multiple non-overlapping local representations and fusing them into a global semantic-rich feature. (2) Based on the information bottleneck theory, we derive Local Information Loss to guarantee the orthogonality of local representations while preserving comprehensive task-relevant information. (3) Further, to fuse the local representations and remove task-irrelevant information, we arrive at a Global Information Loss through the theoretical analysis of mutual information. Empirically, our method achieves state-of-the-art performance on five benchmark datasets. Our code is available at https://github.com/QingyuLiu/Exposing-the-Deception, hoping to inspire researchers.

Quantum neural networks (QNNs) have become a leading paradigm for establishing near-term quantum applications in recent years. The trainability issue of QNNs has garnered extensive attention, spurring demand for a comprehensive analysis of QNNs in order to identify viable solutions. In this work, we propose a perspective that characterizes the trainability of QNNs based on their locality. We prove that the entire variation range of the loss function via adjusting any local quantum gate vanishes exponentially in the number of qubits with a high probability for a broad class of QNNs. This result reveals extra harsh constraints independent of gradients and unifies the restrictions on gradient-based and gradient-free optimizations naturally. We showcase the validity of our results with numerical simulations of representative models and examples. Our findings, as a fundamental property of random quantum circuits, deepen the understanding of the role of locality in QNNs and serve as a guideline for assessing the effectiveness of diverse training strategies for quantum neural networks.

Formal verification of neural networks is a challenging problem due to the complexity and nonlinearity of neural networks. It has been shown that polynomial zonotopes can tightly enclose the output set of a neural network. Unfortunately, the tight enclosure comes with additional complexity in the set representation, thus, rendering subsequent operations expensive to compute, such as computing interval bounds and intersection checking. To address this issue, we present a novel approach to restructure a polynomial zonotope to tightly enclose the original polynomial zonotope while drastically reducing its complexity. The restructuring is achieved by relaxing the exponents of the dependent factors of polynomial zonotopes and finding an appropriate approximation error. We demonstrate the applicability of our approach on output sets of neural networks, where we obtain tighter results in various subsequent operations, such as order reduction, zonotope enclosure, and range bounding.

Length extrapolation has attracted considerable attention recently since it allows transformers to be tested on longer sequences than those used in training. Previous research has shown that this property can be attained by using carefully designed Relative Positional Encodings (RPEs). While these methods perform well on a variety of corpora, the conditions for length extrapolation have yet to be investigated. This paper attempts to determine what types of RPEs allow for length extrapolation through a thorough mathematical and empirical analysis. We discover that a transformer is certain to possess this property as long as the series that corresponds to the RPE's exponential converges. Two practices are derived from the conditions and examined in language modeling tasks on a variety of corpora. As a bonus from the conditions, we derive a new Theoretical Receptive Field (TRF) to measure the receptive field of RPEs without taking any training steps. Extensive experiments are conducted on the Wikitext-103, Books, Github, and WikiBook datasets to demonstrate the viability of our discovered conditions. We also compare TRF to Empirical Receptive Field (ERF) across different models, showing consistently matched trends on these datasets. Code is released at: https://github.com/OpenNLPLab/Rpe.

This paper tackles the problem of efficient and stable video semantic segmentation. While stability has been under-explored, prevalent work in efficient video semantic segmentation uses the keyframe paradigm. They efficiently process videos by only recomputing the low-level features and reusing high-level features computed at selected keyframes. In addition, the reused features stabilize the predictions across frames, thereby improving video consistency. However, dynamic scenes in the video can easily lead to misalignments between reused and recomputed features, which hampers performance. Moreover, relying on feature reuse to improve prediction consistency is brittle; an erroneous alignment of the features can easily lead to unstable predictions. Therefore, the keyframe paradigm exhibits a dilemma between stability and performance. We address this efficiency and stability challenge using a novel yet simple Temporal Feature Correlation (TFC) module. It uses the cosine similarity between two frames’ low-level features to inform the semantic label’s consistency across frames. Specifically, we selectively reuse label-consistent features across frames through linear interpolation and update others through sparse multi-scale deformable attention. As a result, we no longer directly reuse features to improve stability and thus effectively solve feature misalignment. This work provides a significant step towards efficient and stable video semantic segmentation. On the VSPW dataset, our method significantly improves the prediction consistency of image-based methods while being as fast and accurate.

The visual prompts have provided an efficient manner in addressing visual cross-domain problems. Previous works introduce domain prompts to tackle the classification Test-Time Adaptation (TTA) problem by placing image-level prompts on the input and fine-tuning prompts for each target domain. However, since the image-level prompts mask out continuous spatial details in the prompt-allocated region, it will suffer from inaccurate contextual information and limited domain knowledge extraction, particularly when dealing with dense prediction TTA problems. To overcome these challenges, we propose a novel Sparse Visual Domain Prompts (SVDP) approach, which applies minimal trainable parameters (e.g., 0.1%) to pixels across the entire image and reserves more spatial information of the input. To better apply SVDP in extracting domain-specific knowledge, we introduce the Domain Prompt Placement (DPP) method to adaptively allocates trainable parameters of SVDP on the pixels with large distribution shifts. Furthermore, recognizing that each target domain sample exhibits a unique domain shift, we design Domain Prompt Updating (DPU) strategy to optimize prompt parameters differently for each sample, facilitating efficient adaptation to the target domain. Extensive experiments were conducted on widely-used TTA and continual TTA benchmarks, and our proposed method achieves state-of-the-art performance in both semantic segmentation and depth estimation tasks.

Human-Object Interaction (HOI) detection plays a vital role in scene understanding, which aims to predict the HOI triplet in the form of . Existing methods mainly extract multi-modal features (e.g., appearance, object semantics, human pose) and then fuse them together to directly predict HOI triplets. However, most of these methods focus on seeking for self-triplet aggregation, but ignore the potential cross-triplet dependencies, resulting in ambiguity of action prediction. In this work, we propose to explore Self- and Cross-Triplet Correlations (SCTC) for HOI detection. Specifically, we regard each triplet proposal as a graph where Human, Object represent nodes and Action indicates edge, to aggregate self-triplet correlation. Also, we try to explore cross-triplet dependencies by jointly considering instance-level, semantic-level, and layout-level relations. Besides, we leverage the CLIP model to assist our SCTC obtain interaction-aware feature by knowledge distillation, which provides useful action clues for HOI detection. Extensive experiments on HICO-DET and V-COCO datasets verify the effectiveness of our proposed SCTC.

Post-training quantization (PTQ) has emerged as a promising technique for mitigating memory consumption and computational costs in large language models (LLMs). However, a systematic examination of various quantization schemes, model families, and quantization bit precision has been absent from the literature. In this paper, we conduct a comprehensive analysis of these factors by investigating the effects of PTQ on weight-only, activation-only, and weight-and-activation quantization using diverse methods such as round-to-nearest (RTN), GPTQ, ZeroQuant, and their variants. We apply these methods to two distinct model families with parameters ranging from 125M to 176B. Our contributions include: (1) a sensitivity analysis revealing that activation quantization is generally more susceptible to weight quantization, with smaller models often outperforming larger models in terms of activation quantization; (2) an evaluation and comparison of existing PTQ methods to optimize model size reduction while minimizing the impact on accuracy, revealing that none of the current methods can achieve the original model quality for quantization with either INT4-weight or INT4-weight-and-INT8-activation; (3) based on these insights, we propose an optimized method called Low-Rank Compensation (LoRC), which employs low-rank matrices to enhance model quality recovery with a minimal increase in model size.