vgssm_EF15 搭配最佳expand_rate和q取e半值的实验结果

方差实验组

Posted by JoselynZhao on February 27, 2020

说明

  • 运行文件: vrm/vcf03.py
  • 运行命令:
    python3.6 vcf03.py --exp_name vgssm_EF15 --exp_order 0 --EF 15 
    # expand_rate = [1.3,1.2,1.1,1.08,1.06,1.04,1.02]
    # stop_vari_step = len(expand_rate)
    python3.6 vcf03.py --exp_name vgssm_EF15 --exp_order 0 --EF 15  --dataset mars --max_frames 100
    # expand_rate = [1.2,1.1,1.08,1.06,1.04,1.02]
    # stop_vari_step = len(expand_rate)
    
  • 实验结果保存在 vgssm_EF15/0目录下
  • baseline性能如下 image

参数设定

q=1/2e

Duke

|| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | —|—|—|—|—|—|—|—|—| query_rate| 0.65 |0.6|0.55|0.54|0.53|0.52|0.51|-| expand_rate | 1.3| 1.2|1.1|1.08|1.06|1.04|1.02| -|

Mars

|| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | —|—|—|—|—|—|—|—|—| query_rate|0.6|0.55|0.54|0.53|0.52|0.51|-|-| expand_rate | 1.2|1.1|1.08|1.06|1.04|1.02| -|- |

实验结果

duke

image 与baseline性能相比:image

| | mAP | rank-1| rank-5| rank-10 | rank-20| |–|–|–|–|–|–| ours | 59.06 | 68.09 | 81.91 | 84.33 | 87.75 | baseline | 59.21 | 69.08 | 81.19 | - | 88.88| up | -0.15 | +0. 01 |

mars

image 与baseline性能相比:image

  mAP rank-1 rank-5 rank-10 rank-20
ours 30.39 54.03 65.51 71.06 75.1
baseline 29.56 52.32 64.29 - 73.08
up -0.85 +1.71