BiRe-ID: Binary Neural Network for Efficient Person Re-ID

153

Then, we use efficient XNOR and Bit-count operations to replace real-valued operations.

Following [199], the forward process of the BNN is

ai = b^ai−1 ⊙b^wi,

(6.7)

where ⊙represents efficient XNOR and Bit-count operations. Based on XNOR-Net, we

introduce a learnable channel-wise scale factor to modulate the amplitude of real-valued

convolution. Aligned with the Batch Normalization (BN) and activation layers, the 1-bit

convolution is formulated as

b^ai = sign(Φ(αi ◦b^ai−1 ⊙b^wi)).

(6.8)

In KR-GAL, the original output feature ai is first scaled by a channel-wise scale factor

(vector) αi ∈RCi to modulate the amplitude of the real-valued counterparts. It then enters

Φ(·), which represents a composite function built by stacking several layers, e.g., BN layer,

non-linear activation layer, and max pool layer. The output is then binarized to obtain the

binary activations b^ai ∈Bni, using the sign function. sign(·) denotes the sign function that

returns +1 if the input is greater than zeros and −1 otherwise. Then, the 1-bit activation

b^ai can be used for the efficient XNOR and Bit-count of (i + 1)-th layer.

However, the gap in representational capability between wi and b^wi could lead to a

large quantization error. We aim to minimize this performance gap to reduce the quan-

tization error while increasing the binarized kernels’ ability to provide information gains.

Therefore, αi is also used to reconstruct b^wi into wi. This learnable scale factor can lead to

a novel learning process with more precise estimation of convolutional filters by minimizing

a novel adversarial loss. Discriminators D(·) with weights WD are introduced to distinguish

unbinarized kernels wi from reconstructed ones αi ◦b^wi. Therefore, αi and WD are learned

by solving the following optimization problem.

arg

min

wi,b^wi,αi ^max

WD ^{L K}

Adv^{(wi, bwi, αi, WD) + L K}

MSE^{(wi, bwi, αi) ∀i ∈N,}

(6.9)

where L ^K

Adv^{(wi, bwi, αi, WD) is the adversarial loss as}

L ^K

Adv^{(wi, bwi, αi, WD) = log(D(wi; WD)) + log(1 −D(bwi ◦αi; WD)),}

(6.10)

where D(·) consists of several basic blocks, each with a fully connected layer and a

LeakyReLU layer. In addition, we employ discriminators to refine every binarized con-

volution layer during the binarization training process.

Furthermore, LMSE(wi, b^wi, αi) is the kernel loss between the learned real-valued filters

wi and the binarized filters b^wi, which is expressed by MSE as

L ^K

MSE^{(wi, bwi, αi) = λ}

2 ^{||wi −αi ◦bwi||2}

2^,

(6.11)

where MSE is used to balance the gap between real value wi and binarized b^wi. λ is a

balance hyperparameter.

6.2.3

Feature Refining Generative Adversarial Learning (FR-GAL)

We introduce generative adversarial learning (GAL) to refine the low-level characteristic

through self-supervision. We employ the high-level feature with abundant semantic infor-

mation aH ∈RmH to supervise the low-level feature aL ∈RmL, where mH = CH ·WH ·HH

and mL = CL · WL · HL. To keep the channel dimension identical, we first employ a 1 × 1

convolution to reduce CH to CL as

a^∗

H ^{= f(W1×1 ⊗aH),}

(6.12)