xujiashuai/cv_cspnet_video-object-detection_longshortnet
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

cv_cspnet_video-object-detection_longshortnet on Ascend NPU

1. 简介

  • 模型来源: iic/cv_cspnet_video-object-detection_longshortnet
  • 架构: LongShortNet (CSPNet Backbone + SPP + YOLOX Head)
  • 参数量: 27,075,968 (backbone only)
  • 任务: 视频目标检测 (Video Object Detection)
  • 适配状态: SUCCESS
  • 适配时间: 2026-05-18

2. 验证环境

组件版本
torch2.9.0
torch-npu2.9.0.post1
CANN8.5.1
NPUAscend 910B4

3. 精度评测

指标数值
Cosine Similarity0.999999
Max Abs Error0.007388
是否满足要求是(cosine > 0.999)

4. 性能

指标数值
平均延迟6.75 ms
输出形状[1, 1024, 10, 10]

5. 适配说明

从 checkpoint 重建 CSPNet Backbone 结构,包含 stem、dark2-dark5 阶段,每阶段含 ConvBnSiLU 下采样和 C3/CSP 块(dark5 含 SPP 模块)。输入为随机视频帧张量 (1, 12, 320, 320),即 4 帧 × 3 通道。

6. 推理输出证据

Model: iic/cv_cspnet_video-object-detection_longshortnet
Output shape: [1, 1024, 10, 10]
Latency: 6.75ms
Cosine Similarity: 0.999999
Max Abs Error: 0.007388

7. CPU vs NPU 精度对比

指标CPUNPU差异
Cosine Similarity1.00.9999990.0001%
Max Abs Error-0.007388-