Skip to content

Files

Latest commit

 

History

History
26 lines (18 loc) · 1.25 KB

DL.md

File metadata and controls

26 lines (18 loc) · 1.25 KB

ML strategy

Practical:

  • Scale the image down by factor 5 (1280x720 -> 256->144)
  • Only analyze a half to a third of the image (256->144 -> 96 x 54 or 128x64.), centered around the last known ball position
  • When last ball position is unknown, scan the whole image (6 to 9 tiles)
  • Use MobileNet and transfer learning of last layer(s) to detect ball (and maybe blue and red players)
  • Use Depth-Wise Convolution to save around 80% computations/parameters (3x3 conv filter wise followed by 1x1 conv intra-filter)

Wish:

Use a Recurrent neural network, integrating ball movement as derivative. Output is a coordinate (x, y)

Related Work/Links

https://arxiv.org/pdf/1912.05445 -> https://github.com/jac99/FootAndBall https://github.com/DengPingFan/DAVSOD https://paperswithcode.com/task/weakly-supervised-object-detection <- the number one video-object-detection paper https://www.sciencedirect.com/science/article/abs/pii/S123034021830146X https://towardsdatascience.com/detecting-soccer-palyers-and-ball-retinantet-2ab5f997ab2

https://github.com/SteveMacenski/jetson_nano_detection_and_tracking <- mobilenet with

Convert pytorch to TensorRT (6x speed-up)

https://www.learnopencv.com/how-to-convert-a-model-from-pytorch-to-tensorrt-and-speed-up-inference/