164

Applications in Computer Vision






 










 






















FIGURE 6.6

Detailed architecture of 1-bit networks implemented by us. (a) detailed architecture of 1-

bit PointNet. MM denotes matrix multiplication in short; (b) detailed architecture of 1-bit

PointNet++. Cat denotes the concatenation operation; (c) detailed architecture of 1-bit

DGCNN; (d) detailed architecture of the FC unit and the Bi-FC unit used from (a) to (c).

We use 2 BNs in the Bi-FC Unit.

Updating pi: We finally update other parameters pi with wi and αi fixed. δpi is defined

as the gradient of pi. We formulate it as

δpi = ^∂LS

∂pi

(6.63)

pi ←pi −ηδpi.

(6.64)

The above derivations show that POEM is learnable with the BP algorithm. Our POEM

is supervised on the basis of a simple and effective reconstruction loss function. Moreover, we

introduce an efficient Expectation-Maximization algorithm to optimize unbinarized weights,

thus constraining them to formulate a bimodal distribution.

6.3.5

Ablation Study

Hyper-parameter selection: There are hyperparameters λ and τ in Eqs. 6.44 and 6.58

that are related to the reconstruction loss and the EM algorithm. The effect of parameters

λ and τ is evaluated in ModelNet40 for 1-bit PointNet, the architectural details of which

can be found in Fig. 6.6 (a). The Adam optimization algorithm is used during the training

process, with a batch size of 592. Using different values of λ and τ, the performance of

POEM is shown in Table 6.2. In Table 6.2, from left to right lie the overall accuracies (OAs)

with different λ from 1×10⁻³ to 0.

And the OAs with different τ from 1×10⁻² to 0 lie from top to bottom. With a decrease

of λ, the OA increases first and then drops dramatically. The same trend is shown when we