question about decoder with conv?why the less params

[lijx10]

question about decoder with conv?why the less params

[lijx10]

The reasons of using deconv are 2-fold:
1. Less params.
1.1 Take an example that the feature vector is size 1024, and we want to reconstruct 1024 points.

• FC: Assume that the number of neurons are non-increasing. The first layer is at least 1024 -> 31024 (each point has x,y,z coordinates). That is param amount: 102431024. In the following layers, the number of params will be 3102431024.
• deconv: the first layer, the convolution param amount: 102433*1024. But in the following layers, the number of params are decreasing.
• This effect is more obvious if we are reconstructing 5000 points. Each layer of FC takes at least 350003*5000, which is huge. But for deconv, the we have similar number of params as in the case of reconstructing 1024 points.
• That is, deconv is much more scalable.
• Of course, we don't have to assume that the number of neurons are non-increasing for FC. In this case, the number of params depends on how to implement FC/deconv. It is meaning less to make absolute comparison.

2. Deconv has better performance

• In experiments, reconstructing large number points with FC will result in disappointing performance.
• In real objects / scenes, each point is not totally independent to another point. Hence it is not necessary to use FC. Deconv is more suitable.