/TIR-Tracking

HITsz CST2016 Graduation Design - "Fine-grained Feature Based TIR Tracking"

Primary LanguageJupyter Notebook

TIR-Tracking-PixelNet

Graduation Design - Fine-grained Feature Based TIR Tracking

PixelNet embedded SiamRPN++ tracker using ResNet-50 as backbone

In this project, the famous Multi-Head Attention framework is used to build my fine-grained feature extraction network. I combine three Global Context Block with Multi-Head Attention framework and integrate it into SiamRPN++ tracker. The result is satisfying.

But unfortunately when I got this pleasing model, I have done my graduation design and got my Bachelor degree 🙃

模型结构如下:

Nd48eO.png

主要使用框架:PySOT中的SiamRPN++追踪器

课题中增加的代码有:

在离校前得到的结果:

  • SiamRPN++追踪器在RGB图像与TIR图像数据集中的表现

    Dataset Accuracy Robustness Loss EAO
    VOT2019 0.594 0.467 93 0.287
    PTB-TIR 0.404 0.194 51 0.308
  • PixelNet两种嵌入方式的比较

    Model Accuracy Robustness Loss EAO FPS
    PixelNet-2b-a 0.413 0.362 95 0.247 53.6
    PixelNet-2b-b 0.408 0.331 87 0.260 62.4
  • PixelNet(PixelNet-3b)与原模型的比较

    Model Accuracy Robustness Loss EAO FPS
    PixelNet 0.419 0.305 80 0.275 67.9
    Original 0.428 0.350 92 0.264 69.3
  • PixelNet中AttentionBlock的个数的比较

    • 10个epoch时

      AttnBlk No. Accuracy Robustness Loss EAO FPS
      PixelNet-1b 0.378 0.312 82 0.245 72.5
      PixelNet-2b 0.363 0.354 93 0.225 70.3
      PixelNet-3b 0.369 0.312 82 0.240 68.4
      PixelNet-4b 0.357 0.343 90 0.220 66.4
    • 19个epoch时

      AttnBlk No. Accuracy Robustness Loss EAO FPS
      PixelNet-1b 0.418 0.339 89 0.259 72.2
      PixelNet-2b 0.408 0.331 87 0.260 62.4
      PixelNet-3b 0.419 0.305 80 0.275 67.9
      PixelNet-4b 0.399 0.350 92 0.240 64.4