Supplementary Material for Binarized Convolutional Landmark Localizers for Human Pose Estimation and Face Alignment with Limited Resources

This section provides additional details for some of the ablation studies reported in Section 6. Pooling type. In the context of binary networks, and because the output is restricted to 1 and -1, max-pooling might result in outputs full of 1s only. To limit this effect, we placed the activation function before the convolutional layers as proposed in [5, 9]. Additionally, we opted to replace max-pooling with average pooling. However, this leads to slightly worse results (see Table 1). In practice, we found that the use of blocks with pre-activation suffices and that the ratio of 1 and -1 is close to 50% even after max-pooling.