News

Camouflage object detection (COD)is a challenging task that aims to identify targets “seamlessly” concealed within their surrounding environment, presenting a more challenging task compared to traditional object detection[1]. While under the Video camouflage object detection (VCOD), the intrinsic high variations and increased complexity in the scene poses new obstacles for the detection with videos. We proposed a green method, termed GreenCOD, that leverages gradient boosting and deep features extracted from pre-trained Deep Neural Networks (DNNs), efficiently detects the camouflage objects without back-propagation. In this quarter, based on the GreenCOD model, we further move on to explore how to deal with object detection in camouflage videos in a light-weighted, explainable way.

Inspired by the GreenCOD pipeline, our architecture integrates the EfficientNetB4 backbone in the feature extraction module for each frame. Initially, the input frames are first reshaped into a standard size of 672x672x3 and the processed through the 8-block EfficientNetB4 backbone for feature extraction under different size of reception fields. To well consider the information across different reception spatial sizes, all the features are resized and concatenated to form a rich set of features. A hierarchical architecture is implemented in decision learning module.

Then XGBoost is trained based on the initial prediction maps and temporal information. The short term temporal information is considered by extracting the motion among consecutive frames. The motion flow maps are extracted at a higher resolution, then followed by the utilization of neighborhood reconstruction. This approach ensures that each prediction location takes into account the information from a corresponding 4×4 window in the motion map. The initial result show a satisfying result on VCOD problems.

[1] Fan, Deng-Ping, et al. “Camouflaged object detection.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. [...]

Understanding and retrieving information in 3D scenes poses a significant challenge in artificial intelligence (AI) and machine learning (ML), particularly in grasping complex spatial relationships and detailed properties of objects in 3D spaces. While large foundational models such as CLIP [1] have made impressive progress in the 2D domain, their direct application to 3D scene understanding is not straightforward. Therefore, we aim to leverage existing large foundation models to understand 3D scenes and extract 3D information, avoiding building an entirely new 3D model from scratch.
The advancement of large foundation models has inspired recent studies to establish connections among images, text, and 3D data. The work in [2] extracts features from 3D objects with a 3D encoder and aligns them with features extracted by a visual encoder and a text encoder from corresponding rendered 2D images and text, enabling information retrieval of 3D objects from different modalities. However, the model mainly focuses on 3D objects and is limited in understanding complicated 3D spaces like 3D indoor scenes.
[3] Reconstructed compact 3D indoor scenes from multi-view images using DVGO and aligned them with semantic segmentation extracted from corresponding 2D multiview images using the CLIP-LSeg model. A visual reasoning pipeline is applied for reasoning tasks in 3D spaces.
Presently, we are going to develop a ML model that leverages the existing large foundation models to get a better understanding of 3D scenes with lower computational complexity.

Reference:
[1]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. “Learning transferable visual models from natural language supervision,” In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021.
[2] Hegde, Deepti, Jeya Maria Jose Valanarasu, and Vishal [...]

Surface reconstruction from point cloud scans plays a pivotal role in 3D vision and graphics, finding diverse applications in areas such as AR/VR games, cultural heritage preservation, and building information modeling (BIM). This task is inherently challenging due to the ill-posed nature of reconstructing continuous surfaces from discrete points. Moreover, real-world point cloud scans introduce quite a few obstacles such as varying densities and sensor noise. These properties make the problem a long standing one, which keeps driving researchers to look for more effective solutions.
Early research focused on combinatorial methods [1]–[2], which inferred the connectivity between points directly. The mainstream of surface reconstruction adopts an implicit surface approach [3]–[4]. That is, the surface is represented as an unknown continuous function which is solved by the associated partial differential equations (PDEs). Although these methods offer good quality, they often require oriented normals or additional constraints. Recently, people develop deep learning (DL) models [5]–[6] to solve this problem based on a supervised learning framework. DL methods exploits the training data to learn an implicit function representation.
Despite their high reconstruction quality, the generalizability and complexity remain to be challenges for DL models. In scenarios such as point cloud compression, quality assessment, and dynamic point cloud processing, there is a growing need for low-complexity, low-latency surface reconstruction methods. However, existing PDE and DL-based methods tend to sacrifice simplicity for high reconstruction quality, leaving a gap for low-complexity solutions.

We adopt the unsupervised framework, propose a lightweight high-performance method, and name it green point-cloud surface reconstruction (GPSR). It can be categorized to the family of implicit surface reconstruction methods. The main idea lies in building a signed distance field (SDF) through approximated heat diffusion and fine tuning it iteratively. [...]

Saliency detection research predominantly falls within two categories: human eye fixation prediction, which involves the prediction of human gaze locations on images where attention is most concentrated [1], and salient object detection (SOD) [2], which aims to identify salient object regions within an image. Our study specifically focuses on the former saliency prediction, which predicts human gaze from visual stimuli.

Saliency detection constitutes a crucial task in predicting human gaze patterns from visual stimuli. The escalating demand for research in saliency detection is driven by the growing necessity to incorporate such techniques into various computer vision tasks and to understand human visual system. Many existing saliency detection methodologies rely on deep neural networks (DNNs) to achieve good performance. However, the extensive model sizes associated with these approaches impede their integration with other modules or deployment on mobile devices.

To address this need, our study introduces a novel saliency detection method, named “GreenSaliency”, which eschews the use of DNNs while emphasizing a small model footprint and low computational complexity. GreenSaliency comprises two primary steps: 1) multi-layer hybrid feature extraction, and 2) multi-path saliency prediction. Empirical findings demonstrate that GreenSaliency achieves performance levels comparable to certain deep-learning-based (DL-based) methods, while necessitating a considerably smaller model size and significantly reduced computational complexity.

[1] Zhaohui Che, Ali Borji, Guangtao Zhai, Xiongkuo Min, Guodong Guo, and Patrick Le Callet, “How is gaze influenced by image transformations? dataset and model”, IEEE Transactions on Image Processing, 29, 2287–300.

[2] Dingwen Zhang, Junwei Han, Yu Zhang, and Dong Xu, “Synthesizing supervision for learning deep saliency network without human annotation”, IEEE transactions on pattern analysis and machine intelligence, 42(7), 1755–69.

Transfer learning is an approach to extract features from source domain and transfer the knowledge of source domain to the target domain, so as to improve the performance of target learners while relieving the pressure to collect a lot of target-domain data[1]. It has been put into wide applications, such as image classification and text classification.

Domain adaptation on digital number datasets, namely MNIST, USPS and SVHN, is the task of transferring learning models across different data sets, aiming to conduct cross-domain feature alignment and build a generalized model that is able to predict labels among various digital number datasets.
Presently, we have trained a green-learning based transfer learning model between MNIST and USPS. The first step is preprocessing and feature extraction, including feature processing for different dataset in order to make them visually similar, raw Saab feature extraction [2] and LNT feature transformation [3] followed by cosine similarity check to find discriminant features. The second step is joint subspace generation, in which for each label in source domain, k-means with number of clusters as 1, 2, 4, 8 are performed separately in order to generate 10, 20, 40 and 80 subspaces, then assign target features to generated subspaces. The third step is to utilize assigned datas to conduct weakly supervised learning to predict label
for rest target data samples. Our goal is to compare and analyze the performance of our green-learning based transfer learning models with other models. In future, we aim to conduct transfer learning among these three digital number datasets mutually and improve the accuracy by improving cross-domain feature alignment.

[1] Zhuang, Fuzhen, et al. “A comprehensive survey on transfer learning.” Proceedings of the IEEE 109.1 (2020): 43-76.
[2] Y. Chen, M. Rouhsedaghat, [...]