Exploring the Potential of Encoder-free Architectures in 3D LMMs

Encoder-free architectures have been preliminarily explored in the 2D visual domain, yet it remains an open question whether they can be effectively applied to 3D understanding scenarios. In this paper, we present the first comprehensive investigation into the potential of encoder-free architectures to overcome the challenges of encoder-based 3D Large Multimodal Models (LMMs). These challenges include the failure to adapt to varying point cloud resolutions and the point features from the encoder not meeting the semantic needs of Large Language Models (LLMs). We identify key aspects for 3D LMMs to remove the encoder and enable the LLM to assume the role of the 3D encoder: 1) We propose the LLM-embedded Semantic Encoding strategy in the pre-training stage, exploring the effects of various point cloud self-supervised losses. And we present the Hybrid Semantic Loss to extract high-level semantics. 2) We introduce the Hierarchical Geometry Aggregation strategy in the instruction tuning stage. This incorporates inductive bias into the LLM early layers to focus on the local details of the point clouds. To the end, we present the first Encoder-free 3D LMM, ENEL. Our 7B model rivals the current state-of-the-art model, ShapeLLM-13B, achieving 55.0%, 50.92%, and 42.7% on the classification, captioning, and VQA tasks, respectively. Our results demonstrate that the encoder-free architecture is highly promising for replacing encoder-based architectures in the field of 3D understanding. The code is released at https://github.com/Ivan-Tang-3D/ENEL

SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model

In this paper, we claim that spatial understanding is the keypoint in robot manipulation, and propose SpatialVLA to explore effective spatial representations for the robot foundation model. Specifically, we introduce Ego3D Position Encoding to inject 3D information into the input observations of the visual-language-action model, and propose Adaptive Action Grids to represent spatial robot movement actions with adaptive discretized action grids, facilitating learning generalizable and transferrable spatial action knowledge for cross-robot control. SpatialVLA is first pre-trained on top of a vision-language model with 1.1 Million real-world robot episodes, to learn a generalist manipulation policy across multiple robot environments and tasks. After pre-training, SpatialVLA is directly applied to perform numerous tasks in a zero-shot manner. The superior results in both simulation and real-world robots demonstrate its advantage of inferring complex robot motion trajectories and its strong in-domain multi-task generalization ability. We further show the proposed Adaptive Action Grids offer a new and effective way to fine-tune the pre-trained SpatialVLA model for new simulation and real-world setups, where the pre-learned action grids are re-discretized to capture robot-specific spatial action movements of new setups. The superior results from extensive evaluations demonstrate the exceptional in-distribution generalization and out-of-distribution adaptation capability, highlighting the crucial benefit of the proposed spatial-aware representations for generalist robot policy learning. All the details and codes will be open-sourced.

SpatioTemporal Learning for Human Pose Estimation in Sparsely-Labeled Videos

Human pose estimation in videos remains a challenge, largely due to the reliance on extensive manual annotation of large datasets, which is expensive and labor-intensive. Furthermore, existing approaches often struggle to capture long-range temporal dependencies and overlook the complementary relationship between temporal pose heatmaps and visual features. To address these limitations, we introduce STDPose, a novel framework that enhances human pose estimation by learning spatiotemporal dynamics in sparsely-labeled videos. STDPose incorporates two key innovations: 1) A novel Dynamic-Aware Mask to capture long-range motion context, allowing for a nuanced understanding of pose changes. 2) A system for encoding and aggregating spatiotemporal representations and motion dynamics to effectively model spatiotemporal relationships, improving the accuracy and robustness of pose estimation. STDPose establishes a new performance benchmark for both video pose propagation (i.e., propagating pose annotations from labeled frames to unlabeled frames) and pose estimation tasks, across three large-scale evaluation datasets. Additionally, utilizing pseudo-labels generated by pose propagation, STDPose achieves competitive performance with only 26.7% labeled data.

Causal-Inspired Multitask Learning for Video-Based Human Pose Estimation

Video-based human pose estimation has long been a fundamental yet challenging problem in computer vision. Previous studies focus on spatio-temporal modeling through the enhancement of architecture design and optimization strategies. However, they overlook the causal relationships in the joints, leading to models that may be overly tailored and thus estimate poorly to challenging scenes. Therefore, adequate causal reasoning capability, coupled with good interpretability of model, are both indispensable and prerequisite for achieving reliable results. In this paper, we pioneer a causal perspective on pose estimation and introduce a causal-inspired multitask learning framework, consisting of two stages. \textit{In the first stage}, we try to endow the model with causal spatio-temporal modeling ability by introducing two self-supervision auxiliary tasks. Specifically, these auxiliary tasks enable the network to infer challenging keypoints based on observed keypoint information, thereby imbuing causal reasoning capabilities into the model and making it robust to challenging scenes. \textit{In the second stage}, we argue that not all feature tokens contribute equally to pose estimation. Prioritizing causal (keypoint-relevant) tokens is crucial to achieve reliable results, which could improve the interpretability of the model. To this end, we propose a Token Causal Importance Selection module to identify the causal tokens and non-causal tokens (\textit{e.g.}, background and objects). Additionally, non-causal tokens could provide potentially beneficial cues but may be redundant. We further introduce a non-causal tokens clustering module to merge the similar non-causal tokens. Extensive experiments show that our method outperforms state-of-the-art methods on three large-scale benchmark datasets.

Optimizing Human Pose Estimation Through Focused Human and Joint Regions

Human pose estimation has given rise to a broad spectrum of novel and compelling applications, including action recognition, sports analysis, as well as surveillance. However, accurate video pose estimation remains an open challenge. One aspect that has been overlooked so far is that existing methods learn motion clues from all pixels rather than focusing on the target human body, making them easily misled and disrupted by unimportant information such as background changes or movements of other people. Additionally, while the current Transformer-based pose estimation methods has demonstrated impressive performance with global modeling, they struggle with local context perception and precise positional identification. In this paper, we try to tackle these challenges from three aspects: (1) We propose a bilayer Human-Keypoint Mask module that performs coarse-to-fine visual token refinement, which gradually zooms in on the target human body and keypoints while masking out unimportant figure regions. (2) We further introduce a novel deformable cross attention mechanism and a bidirectional separation strategy to adaptively aggregate spatial and temporal motion clues from constrained surrounding contexts. (3) We mathematically formulate the deformable cross attention, constraining that the model focuses solely on the regions centered at the target person body. Empirically, our method achieves state-of-the-art performance on three large-scale benchmark datasets. A remarkable highlight is that our method achieves an 84.8 mean Average Precision (mAP) on the challenging wrist joint, which significantly outperforms the 81.5 mAP achieved by the current state-of-the-art method on the PoseTrack2017 dataset.

A hybrid framework for effective and efficient machine unlearning

Recently machine unlearning (MU) is proposed to remove the imprints of revoked samples from the already trained model parameters, to solve users' privacy concern. Different from the runtime expensive retraining from scratch, there exist two research lines, exact MU and approximate MU with different favorites in terms of accuracy and efficiency. In this paper, we present a novel hybrid strategy on top of them to achieve an overall success. It implements the unlearning operation with an acceptable computation cost, while simultaneously improving the accuracy as much as possible. Specifically, it runs reasonable unlearning techniques by estimating the retraining workloads caused by revocations. If the workload is lightweight, it performs retraining to derive the model parameters consistent with the accurate ones retrained from scratch. Otherwise, it outputs the unlearned model by directly modifying the current parameters, for better efficiency. In particular, to improve the accuracy in the latter case, we propose an optimized version to amend the output model with lightweight runtime penalty. We particularly study the boundary of two approaches in our frameworks to adaptively make the smart selection. Extensive experiments on real datasets validate that our proposals can improve the unlearning efficiency by 1.5$\times$ to 8$\times$ while achieving comparable accuracy.

XKV: Personalized KV Cache Memory Reduction for Long-Context LLM Inference

Recently the generative Large Language Model (LLM) has achieved remarkable success in numerous applications. Notably its inference generates output tokens one-by-one, leading to many redundant computations. The widely-used KV-Cache framework makes a compromise between time and space complexities. However, caching data generates the increasingly growing memory demand, that can quickly exhaust the limited memory capacity of the modern accelerator like GPUs, particularly in long-context inference tasks. Existing studies reduce memory consumption by evicting some of cached data that have less important impact on inference accuracy. But the benefit in practice is far from ideal due to the static cache allocation across different LLM network layers. This paper observes that the layer-specific cached data have very different impacts on accuracy. We quantify this difference, and give experimental and theoretical validation. We accordingly make a formal analysis and shows that customizing the cache size for each layer in a personalized manner can yield a significant memory reduction, while still providing comparable accuracy. We simulate the cache allocation as a combinatorial optimization problem and give a global optimal solution. In particular, we devise a mini- and sampling-based inference over a lightweight variant of the LLM model, so as to quickly capture the difference and then feed it into the personalized algorithms. Extensive experiments on real-world datasets demonstrate that our proposals can reduce KV cache memory consumption by 61.6% on average, improve computational efficiency by 2.1x and then increase the throughput by up to 5.5x.

Open-Vocabulary Octree-Graph for 3D Scene Understanding

Open-vocabulary 3D scene understanding is indispensable for embodied agents. Recent works leverage pretrained vision-language models (VLMs) for object segmentation and project them to point clouds to build 3D maps. Despite progress, a point cloud is a set of unordered coordinates that requires substantial storage space and does not directly convey occupancy information or spatial relation, making existing methods inefficient for downstream tasks, e.g., path planning and complex text-based object retrieval. To address these issues, we propose Octree-Graph, a novel scene representation for open-vocabulary 3D scene understanding. Specifically, a Chronological Group-wise Segment Merging (CGSM) strategy and an Instance Feature Aggregation (IFA) algorithm are first designed to get 3D instances and corresponding semantic features. Subsequently, an adaptive-octree structure is developed that stores semantics and depicts the occupancy of an object adjustably according to its shape. Finally, the Octree-Graph is constructed where each adaptive-octree acts as a graph node, and edges describe the spatial relations among nodes. Extensive experiments on various tasks are conducted on several widely-used datasets, demonstrating the versatility and effectiveness of our method.

Night-to-Day Translation via Illumination Degradation Disentanglement

Night-to-Day translation (Night2Day) aims to achieve day-like vision for nighttime scenes. However, processing night images with complex degradations remains a significant challenge under unpaired conditions. Previous methods that uniformly mitigate these degradations have proven inadequate in simultaneously restoring daytime domain information and preserving underlying semantics. In this paper, we propose \textbf{N2D3} (\textbf{N}ight-to-\textbf{D}ay via \textbf{D}egradation \textbf{D}isentanglement) to identify different degradation patterns in nighttime images. Specifically, our method comprises a degradation disentanglement module and a degradation-aware contrastive learning module. Firstly, we extract physical priors from a photometric model based on Kubelka-Munk theory. Then, guided by these physical priors, we design a disentanglement module to discriminate among different illumination degradation regions. Finally, we introduce the degradation-aware contrastive learning strategy to preserve semantic consistency across distinct degradation regions. Our method is evaluated on two public datasets, demonstrating a significant improvement in visual quality and considerable potential for benefiting downstream tasks.

Towards Flexible and Efficient Diffusion Low Light Enhancer

Diffusion-based Low-Light Image Enhancement (LLIE) has demonstrated significant success in improving the visibility of low-light images. However, the substantial computational burden introduced by the iterative sampling process remains a major concern. Current acceleration methods, whether training-based or training-free, often lead to significant performance degradation. As a result, to achieve an efficient student model with performance comparable to that of existing multi-step teacher model, it is usually necessary to retrain a more capable teacher model. This approach introduces inflexibility, as it requires additional training to enhance the teacher's performance. To address these challenges, we propose \textbf{Re}flectance-aware \textbf{D}iffusion with \textbf{Di}stilled \textbf{T}rajectory (\textbf{ReDDiT}), a step distillation framework specifically designed for LLIE. ReDDiT trains a student model to replicate the teacher's trajectory in fewer steps while also possessing the ability to surpass the teacher's performance. Specifically, we first introduce a trajectory decoder from the teacher model to provide guidance. Subsequently, a reflectance-aware trajectory refinement module is incorporated into the distillation process to enable more deterministic guidance from the teacher model. Our framework achieves comparable performance to previous diffusion-based methods with redundant steps in just 2 steps while establishing new state-of-the-art (SOTA) results with 8 or 4 steps. Comprehensive experimental evaluations on 10 benchmark datasets validate the effectiveness of our method, consistently outperforming existing SOTA methods.