实验

detector lr batch AP_{50} AP_{50}^{tiny} dataset_train dataset_test crop
Faster RCNN 0.01 1x2 73.19 60.63 tinypersonV2RGB RGB-only 640x640
Faster RCNN 0.01 1x2 73.18 60.04 tinypersonV2RGB+Xray RGB-only 640x640
Adap RepPoint(SOTA) 0.04 4x2 75.27 64.56 tinypersonV2RGB RGB-only 640x640
Faster RCNN 0.01 1x2 57.62 56.66 tinypersonV2Xray Xray-only 640x640
FCOS 0.04 4x2 63.57 53.11 tinypersonV2RGB RGB-only 640x640
FCOS 0.04 4x2 30.37 28.45 tinypersonV2infrared infrared-only 640x640
Adap-FCOS 0.04 4x2 71.42 60.59 tinypersonV2RGB RGB-only 640x640
Adap-FCOS 0.04 4x2 60.52 59.99 tinypersonV2infrared infrared-only 640x640
Adap-FCOS 0.04 4x2 68.89 61.80 tinypersonV2RGB+infrared RGB+infrared 640x640
Adap-FCOS 0.04 4x2 74.79 63.18 tinypersonV2RGB+infrared RGB 640x640
Adap-FCOS 0.04 4x2 61.69 60.97 tinypersonV2RGB+infrared infrared 640x640
Adap-FCOS 0.04 4x2 61.78 55.91 separated-model RGB+infrared 640x640

big table

TinyPersonV2

  • cut size: (640, 640)
  • detector:Adap-FCOS
  • lr 0.04 + batch 4x2
train_dataset test_dataset AP_{50} AP_{50}^{tiny}
RGB RGB-only 71.42 60.59
infrared infrared-only 60.52 59.99
RGB+infrared RGB+infrared 68.89 61.80
RGB+infrared RGB-only 74.79 63.18
RGB+infrared infrared-only 61.69 60.97

TinyPersonV3

  • all images - 8548x2
    --- train-val-test 10:1:10
  • detector:Adap-FCOS
train_dataset test_dataset AP_{50} AP_{50}^{tiny}
RGB RGB-only
infrared infrared-only 50.13 41.95
RGB+infrared RGB+infrared 59.80 50.87
RGB+infrared RGB-only 66.90 57.09
RGB+infrared infrared-only 50.63 41.35
Method(train and test) AP_{50} AP_{50}^{tiny}
With overall Attention and Pair Sampler 61.25 52.02
overall Attention 62.28 52.35
overall Attention 61.56 51.92
Pair Sampler only 60.37 51.20
Attention only 59.46 50.79
Random Pair Sampler 60.09 50.40
false Weighted Attention 61.74 51.46
cross attention
separate attention 60.00 51.25
0.5mix attention 58.50 50.16
mix attention
最后编辑于
©著作权归作者所有,转载或内容合作请联系作者
【社区内容提示】社区部分内容疑似由AI辅助生成,浏览时请结合常识与多方信息审慎甄别。
平台声明:文章内容(如有图片或视频亦包括在内)由作者上传并发布,文章内容仅代表作者本人观点,简书系信息发布平台,仅提供信息存储服务。

相关阅读更多精彩内容

  • --- lr0.01/ 2x2gpu lr0.08/4x8 train_datasettest_dataset...
    dasiadi阅读 202评论 0 0
  • 参考资料 ROS stage_ros 与world 模型的讲解The Stage Robot SimulatorM...
    十月石榴2013阅读 1,770评论 0 0
  • 16宿命:用概率思维提高你的胜算 以前的我是风险厌恶者,不喜欢去冒险,但是人生放弃了冒险,也就放弃了无数的可能。 ...
    yichen大刀阅读 8,015评论 0 4
  • 公元:2019年11月28日19时42分农历:二零一九年 十一月 初三日 戌时干支:己亥乙亥己巳甲戌当月节气:立冬...
    石放阅读 7,500评论 0 2
  • 昨天考过了阿里规范,心里舒坦了好多,敲代码也犹如神助。早早完成工作回家喽
    常亚星阅读 3,267评论 0 1

友情链接更多精彩内容