面向统一多任务推理的神经中心视频处理流水线
Neural-Centric Video Processing Pipeline for Unified Multi-Task Inference
用隐式神经表示直接从视频中间层提取任务特征，配合轻量 micro adapter 映射到下游模型表征空间，跳过像素解码与早期特征提取；对做视频多任务推理、压缩域感知或想省掉解码开销的视觉系统是可复用的加速思路。
原文依据 NVP introduces lightweight micro adapters that map these features into the representation space of downstream models, bypassing both decoding and early-stage feature extraction.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论视频正日益成为机器学习系统的输入，而在多种下游任务中反复进行解码与处理主导了计算开销。然而，现有视频流水线仍然低效。H.264 与 H.265 等传统编解码器针对人类感知进行优化，每次查询都需要完整的像素解码；压缩域方法受限于特定编解码器结构，灵活性有限；面向机器的视频编码方法则往往依赖任务专用编码器与彼此独立的表征，且不支持人类可视化。我们提出神经视频流水线（NVP），该框架利用隐式神经表示直接从中间层提取任务相关特征，从而消除像素重建开销。NVP 引入轻量微适配器，将这些特征映射到下游模型的表征空间，绕过解码与早期特征提取。在图像分类、目标检测、动作识别与分割这四项代表性任务上，NVP 将延迟最多降低 89.5%，将推理浮点运算量最多降低 29.9%，同时使用单一统一表征支持多个任务。
英文原文摘要 1232 字符
Videos are increasingly used as inputs to machine learning systems, where repeated decoding and processing across diverse downstream tasks dominate computational cost. However, existing video pipelines remain inefficient. Traditional codecs such as H.264 and H.265 are optimized for human perception and require full pixel decoding for every query, compressed-domain methods are tied to specific codec structures with limited flexibility, and machine-oriented video coding approaches often rely on task-specific encoders and separate representations without supporting human visualization. We propose Neural Video Pipeline (NVP), a framework that leverages implicit neural representations to directly extract task-specific features from intermediate layers, eliminating pixel reconstruction overhead. NVP introduces lightweight micro adapters that map these features into the representation space of downstream models, bypassing both decoding and early-stage feature extraction. Across four representative tasks--image classification, object detection, action recognition, and segmentation--NVP reduces latency by up to 89.5% and inference FLOPs by up to 29.9%, while supporting multiple tasks using a single unified representation.