Escaping the Plateau

You have the architecture and the data, but the network isn't converging. A training loss plateauing at around 0.01 for a coordinate regression task is a classic symptom of Model Collapse—the network has given up and is simply predicting the "average" shape for every input.

1. Learning Rate Discipline

If your learning rate is too high (e.g., $0.001$), the network takes steps that overshoot the target, continuously bouncing around the optimal localized valley without ever dropping in.

The Fix: Manually drop the learning rate to 0.0001 or 0.00001 when the plateau begins. optimizer: tf.train.adam(0.0001)

2. The Huber Loss Strategy

Initially, random 3D perspective tilts produce massive errors. MSE squares these errors, creating gradient spikes that destabilize the weights early on.

The Fix: Use Huber Loss (tf.losses.huberLoss). It acts linearly for large outlier errors (preventing spikes) but acts quadratically like MSE as errors approach zero, ensuring sub-pixel precision in the final phase.

3. Multi-Scale Normalization

Neural networks struggle with disparate scales. Always ensure your input pixels are [0, 1] (divide by 255) and your target coordinates are [0, 1] (divide by canvas dimensions).

4. Output Layer Pureness

Coordinate regression must output negative values if a corner is pulled off-canvas. Never use relu or sigmoid on your final Dense layer. It must stay 'linear'.