@wanghuijiao
2021-04-26T15:22:43.000000Z
字数 3209
阅读 950
实验报告
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/
BP4D_3_fold/id_test_3.txt
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/result/resnet/BP4D_3_fold_3/evaluation_split_3_result_test_mode.json
AU | 论文精度(F1) | 复现(F1) | 复现(Acc) |
---|---|---|---|
1 | 50.2 | 43.8 | 56.0 |
2 | 43.7 | 36.7 | 56.0 |
4 | 57.0 | 53.1 | 81.1 |
6 | 78.5 | 74.2 | 77.2 |
7 | 78.5 | 75.6 | 68.8 |
10 | 82.6 | 80.3 | 75.9 |
12 | 87.0 | 87.6 | 86.3 |
14 | 67.7 | 61.7 | 63.0 |
15 | 49.1 | 48.3 | 83.3 |
17 | 62.4 | 63.7 | 67.6 |
23 | 50.4 | 50.6 | 77.3 |
24 | 49.3 | 53.0 | 84.0 |
Avg | 63.0 | 60.7 | 73.0 |
- 注:论文结果摘取自Table 4, 结果保留一位小数(已四舍五入)
BP4D_3_fold/id_test_3.txt
& ``测试集 | 表情准确率 | 表情误检率 |
---|---|---|
BP4D_3_fold/id_test_3.txt | 37.2% (6840/18393) | 15.6% (1420/9076) |
BP4D_3_fold/id_test_3.txt部分测试 | 40.2% (2868/7141) | 17.1% (1267/7428) |
DISFA | 48.5%(1315/2714) | 21.8%(837/3839) |
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/demo/AU_RCNN_test_demo_10fps.mp4
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/pretrained_model/DISFA_dataset_trained_model/resnet101/3_fold_3_resnet101_linear_model_snapshot.npz
多AU关联检测精度极低:
demo
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/requirements_whj.txt
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/
,虚拟环境为whj_AURCNN
。首先在终端激活虚拟环境,并cd到项目路径中,以下每一项测试的当前路径,若无特殊说明均默认为项目路径:
conda activate whj_AURCNN
cd /ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/
cd ./demo
./run_demo.sh
cd ./demo
./run_demo.sh
/ssd01/wanghuijiao/AU_RCNN/AU_RCNN_init/result/resnet/DISFA_3_fold_3/eval_b50.log
看运行过程输出信息
Line65 key = key_prefix+ "/".join((rgb_img_path.split("/")[-3], rgb_img_path.split("/")[-2],rgb_img_path.split("/")[-1]))
Line93 return None, None, # None
Line143 return new_face, AU_box_dict, # rect
cd ./all_shell_scripts
./evaluate.sh
--cal_acc
值设为0
cd ./demo
./calculate_exp_acc.sh
--cal_acc
值设为1
cd ./demo
./calculate_exp_acc.sh