Learning from a limited number of samples is challenging since the learned model can easily become overfitted based on biased distribution formed by only few training examples. In this paper, we calibrate these few-sample classes transferring statistics with sufficient examples, then an adequate examples be sampled calibrated to expand inputs classifier. We assume every dimension in feature representation follows Gaussian so that mean and variance borrow similar whose are better estimated...

Dataset condensation aims at reducing the network training effort through condensing a cumbersome set into compact synthetic one. State-of-the-art approaches largely rely on learning data by matching gradients between real and batches. Despite intuitive motivation promising results, such gradient-based methods, nature, easily overfit to biased of samples that produce dominant gradients, thus lack global supervision distribution. In this paper, we propose novel scheme Condense dataset...

Hazy images captured under ill-posed scenarios with scattering medium (i.e. haze, fog, or smoke) are contaminated in visibility. Inevitably, these further degraded by noises owing to real-world imaging. Most existing hazy image enhancement methods perform dehazing and denoising stage stage, the undesirable result that estimation error of former has be propagated amplified latter e.g., noise amplification after dehazing. To address this inconsistent degradation, we present an Unsupervised...

The label transition matrix has emerged as a widely accepted method for mitigating noise in machine learning. In recent years, numerous studies have centered on leveraging deep neural networks to estimate the individual instances within context of instance-dependent noise. However, these methods suffer from low search efficiency due large space feasible solutions. Behind this drawback, we explored that real murderer lies invalid class transitions, is, actual probability between certain...

In recent years, growing needs for advanced security and traffic management have significantly heightened the prominence of visible-infrared person re-identification community (VI-ReID), garnering considerable attention. A critical challenge in VI-ReID is performance degradation attributable to label noise, an issue that becomes even more pronounced cross-modal scenarios due increased likelihood data confusion. While previous methods achieved notable successes, they often overlook...

The real-world data tends to be heavily imbalanced and severely skew the data-driven deep neural networks, which makes Long-Tailed Recognition (LTR) a massive challenging task. Existing LTR methods seldom train Vision Transformers (ViTs) with (LT) data, while off-the-shelf pretrain weight of ViTs always leads unfair comparisons. In this paper, we systematically investigate ViTs' performance in propose LiVT from scratch only LT data. With observation that suffer more severe problems, conduct...

Existing methods for single-view 3D object reconstruction directly learn to transform image features into representations. However, these are vulnerable images containing noisy backgrounds and heavy occlusions because the extracted do not contain enough information reconstruct high-quality shapes. Humans routinely use incomplete or visual cues from an retrieve similar shapes their memory shape of object. Inspired by this, we propose a novel method, named Mem3D, that explicitly constructs...

Sheath blight (ShB) causes severe yield loss in rice. Previously, we demonstrated that the sugar will eventually be exported and transporter 11 (SWEET11) mutation significantly improved rice resistance to ShB, but it caused defects seed development. The present study found WRKY36 PIL15 directly activate SWEET11 negatively regulate ShB. Interestingly, interacted with PIL15, activates miR530 a key BR signaling transcription factor WRKY53. AOS2 is an effector protein from Rhizoctonia solani (R....

Online tracking of multiple objects in videos requires strong capacity modeling and matching object appearances. Previous methods for learning appearance embedding mostly rely on instance-level without considering the temporal continuity provided by videos. We design a new instance-to-track objective to learn that compares candidate detection tracks persisted tracker. It enables us not only from labeled with complete tracks, but also unlabeled or partially implement this unified form...

A major gap between few-shot and many-shot learning is the data distribution empirically oserved by model during training. In learning, learned can easily become over-fitted based on biased formed only a few training examples, while ground-truth more accurately uncovered in to learn well-generalized model. this paper, we propose calibrate of these few-sample classes be unbiased alleviate such an over-fitting problem. The calibration achieved transferring statistics from with sufficient...

Recently, facial landmark detection algorithms have achieved remarkable performance on static images. However, these are neither accurate nor stable in motion-blurred videos. The missing of structure information makes it difficult for state-of-the-art to yield good results. In this paper, we propose a framework named FAB that takes advantage consistency the temporal dimension A predictor is proposed predict face structural temporally, which serves as geometry prior. This allows our work...

AdaBoost is a classic ensemble learning algorithm with good classifier performance. In the past, it mainly used weak as base classifier, such KNN. They are simple and easy to train, but essence of impossible get very high classification accuracy. order improve correct rate, this paper introduces based on convolutional neural network, namely adaBoost-CNN, referred ACNN. ACNN design new training method, not only gives weight according error rate in pre-training phase, also dynamically adjusts...

In label-noise learning, estimating the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">transition matrix</i> is a hot topic as matrix plays an important role in building xmlns:xlink="http://www.w3.org/1999/xlink">statistically consistent classifiers</i> . Traditionally, transition from clean labels to noisy (i.e., xmlns:xlink="http://www.w3.org/1999/xlink">clean-label (CLTM)</i> ) has been widely exploited on...

A more realistic object detection paradigm, Open-World Object Detection, has arisen increasing research interests in the community recently. qualified open-world detector can not only identify objects of known categories, but also discover unknown objects, and incrementally learn to categorize them when their annotations progressively arrive. Previous works rely on independent modules recognize categories perform incremental learning, respectively. In this paper, we provide a unified...

This paper proposes a novel framework for multi-label image recognition without any training data, called data-free framework, which uses knowledge of pre-trained Large Language Model (LLM) to learn prompts adapt pretrained Vision-Language (VLM) like CLIP multilabel classification. Through asking LLM by well-designed questions, we acquire comprehensive about characteristics and contexts objects, provides valuable text descriptions learning prompts. Then propose hierarchical prompt method...

Recently, facial landmark detection algorithms have achieved remarkable performance on static images. However, these are neither accurate nor stable in motion-blurred videos. The missing of structure information makes it difficult for state-of-the-art to yield good results. In this paper, we propose a framework named FAB that takes advantage consistency the temporal dimension A predictor is proposed predict face structural temporally, which serves as geometry prior. This allows our work...

A major gap between few-shot and many-shot learning is the data distribution empirically observed by model during training. In learning, learned can easily become over-fitted based on biased formed only a few training examples, while ground-truth more accurately uncovered in to learn well-generalized model. this paper, we propose calibrate of these few-sample classes be unbiased alleviate such an over-fitting problem. The calibration achieved transferring statistics from with sufficient...