=======
Outline of method.
Run it through yolo-ish architecture?
so ((batch_size*seq_size), x, y),
then cut and run through convolutional layers (should use residual blocks to prevent gradient vanishing)
yolo9000
tf.reshape(batch_size, seq_len, x, y)
causal/lstm layers
then backprop the gradients through the causal layers?
Test with longer sequences. Also need pure YOLO detections.
adding start of tf yolo rewrite