Shengkai Sun

Jianfeng Dong, Shengkai Sun, Zhonglin Liu et al.

This paper targets unsupervised skeleton-based action representation learning and proposes a new Hierarchical Contrast (HiCo) framework. Different from the existing contrastive-based solutions that typically represent an input skeleton sequence into instance-level features and perform contrast holistically, our proposed HiCo represents the input into multiple-level features and performs contrast in a hierarchical manner. Specifically, given a human skeleton sequence, we represent it into multiple feature vectors of different granularities from both temporal and spatial domains via sequence-to-sequence (S2S) encoders and unified downsampling modules. Besides, the hierarchical contrast is conducted in terms of four levels: instance level, domain level, clip level, and part level. Moreover, HiCo is orthogonal to the S2S encoder, which allows us to flexibly embrace state-of-the-art S2S encoders. Extensive experiments on four datasets, i.e., NTU-60, NTU-120, PKU-MMD I and II, show that HiCo achieves a new state-of-the-art for unsupervised skeleton-based action representation learning in two downstream tasks including action recognition and retrieval, and its learned action representation is of good transferability. Besides, we also show that our framework is effective for semi-supervised skeleton-based action recognition. Our code is available at https://github.com/HuiGuanLab/HiCo.

Shengkai Sun, Daizong Liu, Jianfeng Dong et al.

Unsupervised pre-training has shown great success in skeleton-based action understanding recently. Existing works typically train separate modality-specific models, then integrate the multi-modal information for action understanding by a late-fusion strategy. Although these approaches have achieved significant performance, they suffer from the complex yet redundant multi-stream model designs, each of which is also limited to the fixed input skeleton modality. To alleviate these issues, in this paper, we propose a Unified Multimodal Unsupervised Representation Learning framework, called UmURL, which exploits an efficient early-fusion strategy to jointly encode the multi-modal features in a single-stream manner. Specifically, instead of designing separate modality-specific optimization processes for uni-modal unsupervised learning, we feed different modality inputs into the same stream with an early-fusion strategy to learn their multi-modal features for reducing model complexity. To ensure that the fused multi-modal features do not exhibit modality bias, i.e., being dominated by a certain modality input, we further propose both intra- and inter-modal consistency learning to guarantee that the multi-modal features contain the complete semantics of each modal via feature decomposition and distinct alignment. In this manner, our framework is able to learn the unified representations of uni-modal or multi-modal skeleton input, which is flexible to different kinds of modality input for robust action understanding in practical cases. Extensive experiments conducted on three large-scale datasets, i.e., NTU-60, NTU-120, and PKU-MMD II, demonstrate that UmURL is highly efficient, possessing the approximate complexity with the uni-modal methods, while achieving new state-of-the-art performance across various downstream task scenarios in skeleton-based action representation learning.

Lipu Zhou, Yaoyun Kang, Junxiang Pang et al.

Dense visual SLAM is a fundamental problem in robotics. Recent advances in 3DGS have demonstrated its potential for dense SLAM. Existing 3DGS frameworks focus on both appearance and geometry modeling. However, scene geometry is typically more critical for SLAM than novel view synthesis because downstream robotic tasks, such as navigation and obstacle avoidance, rely primarily on accurate spatial geometry rather than photorealistic rendering. This observation raises a natural question: Is it feasible for 3DGS to perform 3D reconstruction without scene appearance modeling? Motivated by this, we propose Geometry-only Gaussian Splatting (GeoGS), which directly reconstructs scene geometry, and further present GeoGS-SLAM, a dense visual SLAM system built upon this representation. Specifically, GeoGS retains only spatial parameters to reduce the number of per-primitive parameters by over 80%. In contrast to existing 3DGS methods, GeoGS focuses solely on geometric reconstruction, which significantly reduces the number of Gaussian primitives, accelerates geometric convergence, and enhances robustness to illumination variations. In addition, we present an effective training framework that optimizes the Gaussian primitives via single-view and multi-view geometric and photometric supervision, and speeds up geometry convergence with a local-plane driven initialization that better aligns primitives with local structures. Furthermore, we introduce a map update strategy for loop closure that globally transforms the Gaussian map to align it with the corrected pose estimates, thereby preventing map tearing caused by inconsistent per-viewpoint pose corrections in existing methods. Extensive experiments on synthetic and real-world benchmarks demonstrate that our method outperforms SOTA methods in terms of online mapping efficiency and geometric reconstruction quality.

Zefan Zhang, Kehua Zhu, Shijie Jiang et al.

Video Large Language Models (VideoLLMs) exhibit various types of hallucinations. Existing research has primarily focused on hallucinations involving the presence of events, objects, and scenes in videos, while largely neglecting event relation hallucination. In this paper, we introduce a novel benchmark for evaluating the Video Event Relation Hallucination, named VERHallu. This benchmark focuses on causal, temporal, and subevent relations between events, encompassing three types of tasks: relation classification, question answering, and counterfactual question answering, for a comprehensive evaluation of event relation hallucination. Additionally, it features counterintuitive video scenarios that deviate from typical pretraining distributions, with each sample accompanied by human-annotated candidates covering both vision-language and pure language biases. Our analysis reveals that current state-of-the-art VideoLLMs struggle with dense-event relation reasoning, often relying on prior knowledge due to insufficient use of frame-level cues. Although these models demonstrate strong grounding capabilities for key events, they often overlook the surrounding subevents, leading to an incomplete and inaccurate understanding of event relations. To tackle this, we propose a Key-Frame Propagating (KFP) strategy, which reallocates frame-level attention within intermediate layers to enhance multi-event understanding. Experiments show it effectively mitigates the event relation hallucination without affecting inference speed.