noodles8436 / THE-CROSS

A Smart Traffic Control System for the protection of the socially disadvantaged and rapid transport of emergency vehicles.
GNU General Public License v3.0
0 stars 3 forks source link

Improving the accuracy of Object Detection model. #3

Open noodles8436 opened 3 years ago

noodles8436 commented 3 years ago

The accuracy of the current model is very low at 76%. Therefore, it is necessary to improve the accuracy of the model by adjusting the Hyper-Parameter when learning the model.

Plan : Improving Accuracy of Model

noodles8436 commented 3 years ago

Model : efficientdet_d2_coco17_tpu-32 ( Input Size : 768 x 768 ) num_steps : 300,000 batch_size : 4

=> Evaluation Accuracy : 82% ( +6% increased )

Next Plan : increase num_steps to 600,000

noodles8436 commented 3 years ago

Model : efficientdet_d2_coco17_tpu-32 ( Input Size : 768 x 768 ) num_steps : 600,000 batch_size : 4

=> Evaluation Accuracy : 82% ( +0% increased )

Next Plan : increase efficientdet_d0 model( initial model) & Ensembles Two Models (efficientdet_d0_coco17_tpu-32, efficientdet_d2_coco17_tpu-32)