# experimental_recording **Repository Path**: Muyuen/experimental_recording ## Basic Information - **Project Name**: experimental_recording - **Description**: aisxchaoihcioascio - **Primary Language**: Unknown - **License**: LGPL-3.0 - **Default Branch**: dev - **Homepage**: None - **GVP Project**: No ## Statistics - **Stars**: 1 - **Forks**: 1 - **Created**: 2022-12-17 - **Last Updated**: 2026-05-08 ## Categories & Tags **Categories**: Uncategorized **Tags**: None ## README # 基于PyTorch的手语识别 #### 介绍 以ResNet18和LSTM为编码器,以LSTM为解码器的神经网络模型。数据集采用中科大CSL连续手语数据集。验证集中最高准确率Acc为96.37%、最低错词率WER为5.36%、最低损失值为0.2052。 #### 文件夹创建 在项目datasers/CSL_Continuous/文件夹下创建两个空文件夹分别命名为color和picture。 #### 数据集下载 中科大手语数据集:链接:[SLR数据集](https://pan.baidu.com/s/140JxG7tZ1VBbFkHw3NAeaQ?pwd=quad) 。提取码:**quad** 注:本项目只需下载SLR数据集中的SLR_Dataset/[连续句子]SLR_Dataset/color文件夹内的100个文件,大小为55GB左右,并将其放入本项目SLR-Final-Epoch100/datasets/CSL_Continuous/color文件夹下。 #### 训练步骤 ```git # 下载项目代码 git clone https://gitee.com/zoujc2001/gesture-recognition.git # 对数据集视频进行关键帧提取,可能需要耗时2小时 python cut_frame_more.py # 开始训练 python CSL_Continuous_Seq2Seq.py ``` #### 训练结果 | Dataset | Sentences | Samples | Best Test Wer | Best Test Loss | Acc | | -------------- | --------- | ------- | ------------- | -------------- | ------ | | CSL_Continuous | 100 | 25,000 | 5.36% | 0.2052 | 96.37% | ![image-20221126004347064](README.assets/image-20221126004347064.png) ![image-20221126004418097](README.assets/image-20221126004418097.png) ![image-20221126004445500](README.assets/image-20221126004445500.png)