face-alignment | 7000FPS face alignment | Graphics library

What usually consumes most of the memory are the activation maps (and gradients, when training). I am not familiar with this particular model and implementation, but I would say that you are using a "fake" binary network; by fake I mean they still use floating-point numbers to represent the binary values since most users are going to use their code on GPUs that do not fully support real binary operations. The authors even write in Section 5:

Performance. In theory, by replacing all floating-point multiplications with bitwise XOR and making use of the SWAR (Single instruction, multiple data within a register) [5], [6], the number of operations can be reduced up to 32x when compared against the multiplication-based convolution. However, in our tests, we observed speedups of up to 3.5x, when compared against cuBLAS, for matrix multiplications, a result being in accordance with those reported in [6]. We note that we did not conduct experiments on CPUs. However, given the fact that we used the same method for binarization as in [5], similar improvements in terms of speed, of the order of 58x, are to be expected: as the realvalued network takes 0.67 seconds to do a forward pass on a i7-3820 using a single core, a speedup close to x58 will allow the system to run in real-time. In terms of memory compression, by removing the biases, which have minimum impact (or no impact at all) on performance, and by grouping and storing every 32 weights in one variable, we can achieve a compression rate of 39x when compared against the single precision counterpart of Torch.

In this context, a small model (w.r.t. number of parameters or model size in MiB) does not necessarily mean low memory footprint. It is likely that all this memory is being used to store the activation maps in single- or double-precision.