Trained with 1 GPU H800 Server from AutoDL on 2025.2.3 BJS with Pytroch and converted to .h5 format at the same time.
Basic Model uses CNN with accuracy of 75% on test data (80.7 MB)
V1 Engine uses CNN with accuracy of 87% on test data (72.1 MB)
V2 Engine uses ViT with accuracy of at most 40% Keyboard Interrupted 2025.2.3 15:57:37 BJS
V3 Engine uses Hybrid Model( Combination of Convolutional layers and a Multi-Layer Perceptron (MLP)) with accuracy 68.65% on test data. (34.3 MB)
Trained 2025.2.4 BJS with H800 V4 Engine based of V1 but improve with: More Convolutional Layers. Bottleneck Blocks: We can use bottleneck blocks (1x1 conv before and after 3x3 conv) to reduce computation, and increase depth. Residual Connections: Implement residual connections to ease training in the very deep network and to help avoid vanishing gradients. Increased Filters: Use more filters in the layers to increase the learning capacity. Accuracy 89.39% on test data.
E1 Engine : The technology used in this solution combines EfficientNet-B0 as the base model, enhanced by knowledge distillation from a ResNet-34 teacher model to improve accuracy, and quantization to reduce the model size. After training and optimization, the final quantized model achieves a compact size of 16.6 MB, making it highly efficient for deployment. On the test dataset, the model delivers a strong final accuracy of 93.78%, demonstrating its effectiveness in jersey number detection while meeting strict size constraints.