1 views collected around this technical thread.
This article introduces DeepLabV2, explains its challenges, architectural enhancements such as the ASPP module, backbone modifications, poly learning‑rate policy, and presents experimental comparisons on several benchmark datasets, providing a concise yet comprehensive overview for computer‑vision practitioners.
This article explains the two main challenges of applying deep convolutional neural networks to semantic segmentation—signal down‑sampling and loss of spatial precision—and describes how the DeepLabV1 architecture, using dilated convolutions, large‑field‑of‑view modules, fully‑connected CRF and multi‑scale fusion, addresses these issues while achieving faster, more accurate segmentation results.
This tutorial walks through the complete FCN pipeline for semantic segmentation, covering VOC dataset loading, data augmentation, collate functions, model construction, training loops, loss computation with cross‑entropy (including ignore‑index handling), and inference, while providing full PyTorch code snippets for each step.
This article provides a comprehensive, beginner‑friendly overview of semantic segmentation, focusing on the pioneering Fully Convolutional Network (FCN) architecture, its variants (FCN‑32s, FCN‑16s, FCN‑8s), underlying concepts, loss computation, and practical tips for working with the VOC dataset.
To enable real‑time bullet‑comment passthrough on Youku’s mobile app, the team built a million‑scale portrait dataset and designed the AirSegNet series—CPU, GPU, and server variants—using VGG‑style nets, edge‑aware losses, and hybrid CPU‑GPU inference, achieving 0.98 IoU and sub‑15 ms latency on most devices.
The article presents a comprehensive AI solution for video danmu (bullet‑screen) occlusion prevention that leverages human portrait semantic segmentation, describes dataset construction with full‑ and semi‑supervised labeling, details the encoder‑decoder model with context attention, outlines post‑processing for spatial and temporal stability, and explains deployment on cloud and edge using YKit, KwaiNN and TensorRT.
The article presents Semantic Image Matting, a novel approach that incorporates 20 semantic Alpha pattern categories into the matting pipeline via semantic Trimap, region‑based classifiers, multi‑class discriminators, and learnable gradient loss, achieving state‑of‑the‑art results on multiple benchmarks.
Gaode leverages high‑resolution satellite imagery as an active discovery tool for map updates, extracting road, region and building elements through advanced semantic segmentation networks (U‑Net, ASPP, attention, non‑local) and instance‑segmentation pipelines, to accelerate accurate road‑network and building‑block data refreshes while addressing future scalability challenges.
Gaode Maps has progressed image segmentation from early heuristic region splitting to modern deep‑learning pipelines—leveraging FCNs, multi‑task networks, Mask R‑CNN, and specialized losses—to achieve centimeter‑level, instance‑aware mapping of roads, signs, and small objects while pursuing lighter, real‑time models.
The DFS face-detection algorithm jointly created by Didi AI Labs and Beijing University's PRIS team secured five first-place and one second-place results on the WIDER FACE benchmark, achieving 96.3% (Easy), 95.4% (Medium) and 90.7% (Hard) AP by leveraging a Feature Fusion Pyramid and semantic-segmentation supervision, and is already deployed in Didi's driver-identity verification and in-vehicle privacy systems.
This article provides an extensive review of deep learning techniques applied to computer vision, covering the evolution of CNN architectures, image and video processing tasks, 2.5‑D and 3‑D reconstruction, object detection, segmentation, tracking, SLAM, and various practical applications such as AR, content retrieval, and autonomous driving.
iQIYI’s bullet‑screen masking employs a DeepLabv3+‑based two‑class semantic segmentation pipeline, preceded by a close‑up detector and followed by morphological refinement, trained on a custom annotated dataset that raises IoU to 93.6 %, processes hour‑long videos in under an hour, and is slated for future upgrades to instance and panoptic segmentation for finer‑grained masking.
This article traces the evolution of computer vision from its 1960s origins, explains the challenges of image classification and semantic segmentation, and introduces pixel-level decision forest algorithms as an efficient solution for large‑scale pixel classification tasks.