Highlight

Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

Introduces PVT, a convolution-free pyramid Transformer backbone for dense prediction tasks like detection and segmentation.

The Pyramid Vision Transformer (PVT) is a convolution-free backbone for dense prediction, unlike ViT which targets image classification. A progressive shrinking pyramid lets PVT produce high-resolution outputs while reducing computation on large feature maps, and it inherits advantages of both CNNs and Transformers as a unified backbone that directly replaces CNN backbones. Across object detection, instance and semantic segmentation it boosts performance; PVT+RetinaNet reaches 40.4 AP on COCO, surpassing ResNet50+RetinaNet's 36.3 AP by 4.1 points.

Based on: Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

CNN Features Off-the-Shelf: An Astounding Baseline for Recognition

Shows off-the-shelf CNN features from OverFeat, with a linear SVM, beat tuned state-of-the-art systems across many recognition tasks.

Generic descriptors extracted from convolutional networks are remarkably powerful. Using features from the publicly available OverFeat network, trained for object classification on ILSVRC13, as a generic image representation, the paper addresses tasks moving further from the original: object and scene recognition, fine-grained recognition, attribute detection, and image retrieval. A simple linear SVM (or L2 distance for retrieval) on a 4096-dimensional feature consistently beats highly tuned state-of-the-art systems, suggesting deep CNN features should be the default choice.

Based on: CNN Features Off-the-Shelf: An Astounding Baseline for Recognition · 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops

Curated by Aramai Editorial

Read summary →
Highlight

Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)

Surveys explainable AI (XAI), reviewing approaches, trends, and research directions for making black-box AI systems transparent.

This survey addresses the lack of transparency in AI systems, whose black-box nature enables powerful predictions that cannot be directly explained. It frames explainable AI (XAI) as a field holding substantial promise for improving trust and transparency, and as essential for AI to keep making progress without disruption. Through the lens of the literature, it provides an entry point for researchers and practitioners, reviewing existing approaches, discussing trends, and presenting major research trajectories.

Based on: Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI) · IEEE Access

Curated by Aramai Editorial

Read summary →
Highlight

OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks

Introduces OverFeat, an integrated ConvNet framework performing classification, localization, and detection with a single shared network.

OverFeat is an integrated framework that uses convolutional networks for classification, localization, and detection. It efficiently implements a multiscale, sliding-window approach within a ConvNet and introduces a localization method that learns to predict object boundaries, accumulating bounding boxes to increase detection confidence. A single shared network learns all three tasks simultaneously. The framework won the ILSVRC2013 localization task, gave competitive detection and classification results, and set a new state of the art for detection.

Based on: OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Qwen2.5-VL Technical Report

Presents Qwen2.5-VL, a vision-language model with strong object localization, document parsing, and long-video understanding.

Qwen2.5-VL is the flagship Qwen vision-language model, advancing visual recognition, precise object localization via bounding boxes or points, robust document and structured-data parsing, and long-video comprehension. It adds dynamic resolution processing and absolute time encoding to handle varying image sizes and hours-long videos with second-level event localization, using a natively trained dynamic-resolution ViT with window attention. Its 72B version matches models like GPT-4o and Claude 3.5 Sonnet while retaining strong language skills.

Based on: Qwen2.5-VL Technical Report · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

Deep High-Resolution Representation Learning for Human Pose Estimation

Proposes HRNet, which maintains high-resolution representations throughout the network for more precise human pose estimation.

The paper addresses human pose estimation by learning reliable high-resolution representations. Unlike methods that recover high-resolution features from low-resolution ones, the proposed network maintains high resolution throughout by connecting multi-resolution subnetworks in parallel and repeatedly fusing them across scales. This yields richer representations and more accurate, spatially precise keypoint heatmaps. It achieves superior results on the COCO keypoint, MPII, and PoseTrack benchmarks, with code released publicly.

Based on: Deep High-Resolution Representation Learning for Human Pose Estimation · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Understanding deep learning requires rethinking generalization

Shows that large neural networks can perfectly fit random labels, challenging conventional explanations of why deep nets generalize.

The paper questions why large neural networks generalize well despite their size. Through systematic experiments, it shows state-of-the-art convolutional networks trained with stochastic gradient methods can easily fit random labels, and even unstructured random noise, with explicit regularization having little effect. A theoretical construction shows simple depth-two networks already achieve perfect finite-sample expressivity once parameters exceed data points, indicating traditional explanations of generalization are insufficient.

Based on: Understanding deep learning requires rethinking generalization · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

Wasserstein GAN

Introduces WGAN, an alternative GAN training algorithm that improves training stability and eliminates mode collapse.

The paper introduces WGAN, an alternative to conventional GAN training. It improves the stability of learning, removes issues such as mode collapse, and yields meaningful learning curves useful for debugging and hyperparameter searches. The authors show that the corresponding optimization problem is sound and provide extensive theoretical work linking it to other distances between probability distributions.

Based on: Wasserstein GAN · arXiv.org

Curated by Aramai Editorial

Read summary →
Highlight

What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?

Presents a Bayesian deep learning framework combining aleatoric and epistemic uncertainty, improving semantic segmentation and depth regression.

Two uncertainties matter: aleatoric, the noise inherent in observations, and epistemic, model uncertainty that more data can explain away. Epistemic uncertainty is hard to model in vision, but new Bayesian deep learning tools make it tractable. They present a framework combining input-dependent aleatoric and epistemic uncertainty, studied on semantic segmentation and depth regression. Their explicit formulation yields new loss functions interpretable as learned attenuation, making training robust to noisy data and setting new state-of-the-art on these benchmarks.

Based on: What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? · Neural Information Processing Systems

Curated by Aramai Editorial

Read summary →
Highlight

Learning Transferable Architectures for Scalable Image Recognition

Learns transferable CNN cells by searching a building block on CIFAR-10 (the NASNet search space) and stacking it for ImageNet-scale recognition.

Because designing image classification networks demands heavy engineering, this work learns architectures directly from data. Since searching a large dataset is costly, it searches a convolutional cell on small CIFAR-10 and transfers it to ImageNet by stacking copies, enabled by a new transferable 'NASNet search space' and a ScheduledDropPath regularizer. The best cell gives 2.4% error on CIFAR-10 and, stacked, 82.7% top-1 on ImageNet, 1.2% above the best human-designed models with 9 billion fewer FLOPS. Its features also transfer to COCO detection at 43.1% mAP.

Based on: Learning Transferable Architectures for Scalable Image Recognition · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Striving for Simplicity: The All Convolutional Net

Shows max-pooling can be replaced by strided convolutions, proposing an all-convolutional network competitive on CIFAR and ImageNet.

Most modern CNNs for object recognition alternate convolution and max-pooling layers followed by a few fully connected layers. The authors re-evaluate small-image object recognition, questioning whether each component is necessary. They find max-pooling can be replaced by a convolutional layer with increased stride without losing accuracy on several benchmarks. They then propose an architecture made solely of convolutional layers that yields competitive or state-of-the-art results on CIFAR-10, CIFAR-100, and ImageNet, plus a new deconvolution-based feature visualization.

Based on: Striving for Simplicity: The All Convolutional Net · International Conference on Learning Representations

Curated by Aramai Editorial

Read summary →
Highlight

FCOS: Fully Convolutional One-Stage Object Detection

Proposes FCOS, an anchor-free, proposal-free one-stage object detector that predicts objects per-pixel, analogous to semantic segmentation.

FCOS is a fully convolutional one-stage object detector solving detection by per-pixel prediction like semantic segmentation. Unlike leading detectors such as RetinaNet, SSD, YOLOv3, and Faster R-CNN, which depend on pre-defined anchor boxes, FCOS is anchor-box and proposal free, avoiding complex overlap computation and sensitive anchor hyperparameters. Using only non-maximum suppression, FCOS with a ResNeXt-64x4d-101 backbone reaches 44.7% AP with single-model, single-scale testing, surpassing prior one-stage detectors while being simpler and more flexible.

Based on: FCOS: Fully Convolutional One-Stage Object Detection · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →