Let (B) be batch size, (H\times W) input resolution, (C_{in}, C_{out}) channel counts, (K) convolution-kernel width, and (d) flattened feature dimension.
| Operation | Best | Average | Worst | Space |
|---|---|---|---|---|
| Pixel-string parsing (FERDataset) | (\Theta(HW)) | (\Theta(HW)) | (\Theta(HW)) | (\Theta(HW)) per sample |
| 2-D convolution | (\Theta(BHWC_{in}C_{out}K^2)) | same | same | activations (\Theta(BHWC_{out})) |
| Max pooling | (\Theta(BHWC)) | same | same | output activations |
| Dense layer | (\Theta(BdC_{out})) | same | same | parameters plus batch activations |
| Softmax over seven logits | (\Theta(7B)) | same | same | (\Theta(7B)) |
| Reference confusion matrix | (\Theta(N+K^2)) | same | same | (\Theta(K^2)) |
EmotionCNN contains two 3x3 convolution blocks, each followed by 2x2 pooling, then dense layers of dimensions 64x12x12 to 128 to 7. Convolutional activation memory dominates small-batch inference, while the first dense layer has a substantial fixed parameter block. The reference metric implementation is (\Theta(N+K^2)) for (K=7).
Asymptotic notation is not a latency measurement. Hardware, decoding, transfer, batch size, kernel implementation, and model loading can dominate runtime. Follow docs/BENCHMARKING.md and record hardware and model checksum before making empirical claims.