Event camera, as an emerging biologically-inspired vision sensor for capturing motion dynamics, presents new potential for 3D human pose tracking, or video-based 3D human pose estimation. However, existing works in pose tracking either require the presence of additional gray-scale images to establish a solid starting pose, or ignore the temporal dependencies all together by collapsing segments of event streams to form static image frames. Meanwhile, although the effectiveness of Artificial Neural Networks (ANNs, a.k.a. dense deep learning) has been showcased in many event-based tasks, the use of ANNs tends to neglect the fact that compared to the dense frame-based image sequences, the occurrence of events from an event camera is spatiotemporally much sparser. Motivated by the above mentioned issues, we present in this paper a dedicated end-to-end \textit{sparse deep learning} approach for event-based pose tracking: 1) to our knowledge this is the first time that 3D human pose tracking is obtained from events only, thus eliminating the need of accessing to any frame-based images as part of input; 2) our approach is based entirely upon the framework of Spiking Neural Networks (SNNs), which consists of Spike-Element-Wise (SEW) ResNet and our proposed spiking spatiotemporal transformer; 3) a large-scale synthetic dataset is constructed that features a broad and diverse set of annotated 3D human motions, as well as longer hours of event stream data, named SynEventHPD. Empirical experiments demonstrate the superiority of our approach in both performance and efficiency measures. For example, with comparable performance to the state-of-the-art ANNs counterparts, our approach achieves a computation reduction of 20\% in FLOPS. Our implementation is made available at https://github.com/JimmyZou/HumanPoseTracking_SNN and dataset will be released upon paper acceptance.
Continual learning (CL) is one of the most promising trends in recent machine learning research. Its goal is to go beyond classical assumptions in machine learning and develop models and learning strategies that present high robustness in dynamic environments. The landscape of CL research is fragmented into several learning evaluation protocols, comprising different learning tasks, datasets, and evaluation metrics. Additionally, the benchmarks adopted so far are still distant from the complexity of real-world scenarios, and are usually tailored to highlight capabilities specific to certain strategies. In such a landscape, it is hard to objectively assess strategies. In this work, we fill this gap for CL on image data by introducing two novel CL benchmarks that involve multiple heterogeneous tasks from six image datasets, with varying levels of complexity and quality. Our aim is to fairly evaluate current state-of-the-art CL strategies on a common ground that is closer to complex real-world scenarios. We additionally structure our benchmarks so that tasks are presented in increasing and decreasing order of complexity -- according to a curriculum -- in order to evaluate if current CL models are able to exploit structure across tasks. We devote particular emphasis to providing the CL community with a rigorous and reproducible evaluation protocol for measuring the ability of a model to generalize and not to forget while learning. Furthermore, we provide an extensive experimental evaluation showing that popular CL strategies, when challenged with our benchmarks, yield sub-par performance, high levels of forgetting, and present a limited ability to effectively leverage curriculum task ordering. We believe that these results highlight the need for rigorous comparisons in future CL works as well as pave the way to design new CL strategies that are able to deal with more complex scenarios.
The task of dimensionality reduction and visualization of high-dimensional datasets remains a challenging problem since long. Modern high-throughput technologies produce newer high-dimensional datasets having multiple views with relatively new data types. Visualization of these datasets require proper methodology that can uncover hidden patterns in the data without affecting the local and global structures within the data. To this end, however, very few such methodology exist, which can realise this task. In this work, we have introduced a novel unsupervised deep neural network model, called NeuroDAVIS, for data visualization. NeuroDAVIS is capable of extracting important features from the data, without assuming any data distribution, and visualize effectively in lower dimension. It has been shown theoritically that neighbourhood relationship of the data in high dimension remains preserved in lower dimension. The performance of NeuroDAVIS has been evaluated on a wide variety of synthetic and real high-dimensional datasets including numeric, textual, image and biological data. NeuroDAVIS has been highly competitive against both t-Distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP) with respect to visualization quality, and preservation of data size, shape, and both local and global structure. It has outperformed Fast interpolation-based t-SNE (Fit-SNE), a variant of t-SNE, for most of the high-dimensional datasets as well. For the biological datasets, besides t-SNE, UMAP and Fit-SNE, NeuroDAVIS has also performed well compared to other state-of-the-art algorithms, like Potential of Heat-diffusion for Affinity-based Trajectory Embedding (PHATE) and the siamese neural network-based method, called IVIS. Downstream classification and clustering analyses have also revealed favourable results for NeuroDAVIS-generated embeddings.
Human-object interaction (HOI) detection aims to extract interacting human-object pairs and their interaction categories from a given natural image. Even though the labeling effort required for building HOI detection datasets is inherently more extensive than for many other computer vision tasks, weakly-supervised directions in this area have not been sufficiently explored due to the difficulty of learning human-object interactions with weak supervision, rooted in the combinatorial nature of interactions over the object and predicate space. In this paper, we tackle HOI detection with the weakest supervision setting in the literature, using only image-level interaction labels, with the help of a pretrained vision-language model (VLM) and a large language model (LLM). We first propose an approach to prune non-interacting human and object proposals to increase the quality of positive pairs within the bag, exploiting the grounding capability of the vision-language model. Second, we use a large language model to query which interactions are possible between a human and a given object category, in order to force the model not to put emphasis on unlikely interactions. Lastly, we use an auxiliary weakly-supervised preposition prediction task to make our model explicitly reason about space. Extensive experiments and ablations show that all of our contributions increase HOI detection performance.
Human brains respond to semantic features of presented stimuli with different neurons. It is then curious whether modern deep neural networks admit a similar behavior pattern. Specifically, this paper finds a small cluster of neurons in a diffusion model corresponding to a particular subject. We call those neurons the concept neurons. They can be identified by statistics of network gradients to a stimulation connected with the given subject. The concept neurons demonstrate magnetic properties in interpreting and manipulating generation results. Shutting them can directly yield the related subject contextualized in different scenes. Concatenating multiple clusters of concept neurons can vividly generate all related concepts in a single image. A few steps of further fine-tuning can enhance the multi-concept capability, which may be the first to manage to generate up to four different subjects in a single image. For large-scale applications, the concept neurons are environmentally friendly as we only need to store a sparse cluster of int index instead of dense float32 values of the parameters, which reduces storage consumption by 90\% compared with previous subject-driven generation methods. Extensive qualitative and quantitative studies on diverse scenarios show the superiority of our method in interpreting and manipulating diffusion models.
Tons of images and videos are fed into machines for visual recognition all the time. Like human vision system (HVS), machine vision system (MVS) is sensitive to image quality, as quality degradation leads to information loss and recognition failure. In recent years, MVS-targeted image processing, particularly image and video compression, has emerged. However, existing methods only target an individual machine rather than the general machine community, thus cannot satisfy every type of machine. Moreover, the MVS characteristics are not well leveraged, which limits compression efficiency. In this paper, we introduce a new concept, Satisfied Machine Ratio (SMR), to address these issues. SMR statistically measures the image quality from the machine's perspective by collecting and combining satisfaction scores from a large quantity and variety of machine subjects, where such scores are obtained with MVS characteristics considered properly. We create the first large-scale SMR dataset that contains over 22 million annotated images for SMR studies. Furthermore, a deep learning-based model is proposed to predict the SMR for any given compressed image or video frame. Extensive experiments show that using the SMR model can significantly improve the performance of machine recognition-oriented image and video compression. And the SMR model generalizes well to unseen machines, compression frameworks, and datasets.
Continual learning (CL) can help pre-trained vision-language models efficiently adapt to new or under-trained data distributions without re-training. Nevertheless, during the continual training of the Contrastive Language-Image Pre-training (CLIP) model, we observe that the model's zero-shot transfer ability significantly degrades due to catastrophic forgetting. Existing CL methods can mitigate forgetting by replaying previous data. However, since the CLIP dataset is private, replay methods cannot access the pre-training dataset. In addition, replaying data of previously learned downstream tasks can enhance their performance but comes at the cost of sacrificing zero-shot performance. To address this challenge, we propose a novel method ZSCL to prevent zero-shot transfer degradation in the continual learning of vision-language models in both feature and parameter space. In the feature space, a reference dataset is introduced for distillation between the current and initial models. The reference dataset should have semantic diversity but no need to be labeled, seen in pre-training, or matched image-text pairs. In parameter space, we prevent a large parameter shift by averaging weights during the training. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate different methods, where tasks are from various domains instead of class-separated in a single dataset. Our method outperforms other methods in the traditional class-incremental learning setting and the MTIL by 9.7% average score. Our code locates at https://github.com/Thunderbeee/ZSCL.
Accurately estimating the 6D pose of objects is crucial for many applications, such as robotic grasping, autonomous driving, and augmented reality. However, this task becomes more challenging in poor lighting conditions or when dealing with textureless objects. To address this issue, depth images are becoming an increasingly popular choice due to their invariance to a scene's appearance and the implicit incorporation of essential geometric characteristics. However, fully leveraging depth information to improve the performance of pose estimation remains a difficult and under-investigated problem. To tackle this challenge, we propose a novel framework called SwinDePose, that uses only geometric information from depth images to achieve accurate 6D pose estimation. SwinDePose first calculates the angles between each normal vector defined in a depth image and the three coordinate axes in the camera coordinate system. The resulting angles are then formed into an image, which is encoded using Swin Transformer. Additionally, we apply RandLA-Net to learn the representations from point clouds. The resulting image and point clouds embeddings are concatenated and fed into a semantic segmentation module and a 3D keypoints localization module. Finally, we estimate 6D poses using a least-square fitting approach based on the target object's predicted semantic mask and 3D keypoints. In experiments on the LineMod and Occlusion LineMod datasets, SwinDePose outperforms existing state-of-the-art methods for 6D object pose estimation using depth images. This demonstrates the effectiveness of our approach and highlights its potential for improving performance in real-world scenarios. Our code is at https://github.com/zhujunli1993/SwinDePose.
We present a new approach for learning Mobile Realistic Fullbody (MoRF) avatars. MoRF avatars can be rendered in real-time on mobile phones, have high realism, and can be learned from monocular videos. As in previous works, we use a combination of neural textures and the mesh-based body geometry modeling SMPL-X. We improve on prior work, by learning per-frame warping fields in the neural texture space, allowing to better align the training signal between different frames. We also apply existing SMPL-X fitting procedure refinements for videos to improve overall avatar quality. In the comparisons to other monocular video-based avatar systems, MoRF avatars achieve higher image sharpness and temporal consistency. Participants of our user study also preferred avatars generated by MoRF.
Multiple Instance Learning (MIL) has become the predominant approach for classification tasks on gigapixel histopathology whole slide images (WSIs). Within the MIL framework, single WSIs (bags) are decomposed into patches (instances), with only WSI-level annotation available. Recent MIL approaches produce highly informative bag level representations by utilizing the transformer architecture's ability to model the dependencies between instances. However, when applied to high magnification datasets, problems emerge due to the large number of instances and the weak supervisory learning signal. To address this problem, we propose to additionally train transformers with a novel Bag Embedding Loss (BEL). BEL forces the model to learn a discriminative bag-level representation by minimizing the distance between bag embeddings of the same class and maximizing the distance between different classes. We evaluate BEL with the Transformer architecture TransMIL on two publicly available histopathology datasets, BRACS and CAMELYON17. We show that with BEL, TransMIL outperforms the baseline models on both datasets, thus contributing to the clinically highly relevant AI-based tumor classification of histological patient material.