Jungduri
commented
1 year ago 전체 Forward
# FIT = Transformer(injection=True) # feature injecting transformer
# SET = Transformer(injection=False) # self enhancing transformer
def forward(self, inputs, targets, meta_info, mode):
p_feats, s_feats = self.backbone(inputs['img']) # primary, secondary feats
feats = self.FIT(s_feats, p_feats)
feats = self.SET(feats, feats)
if mode == 'train':
gt_mano_params = torch.cat([targets['mano_pose'], targets['mano_shape']], dim=1)
else:
gt_mano_params = None
pred_mano_results, gt_mano_results, preds_joints_img = self.regressor(feats, gt_mano_params)
if mode == 'train':
# loss functions
loss = {}
loss['mano_verts'] = cfg.lambda_mano_verts * F.mse_loss(pred_mano_results['verts3d'], gt_mano_results['verts3d'])
loss['mano_joints'] = cfg.lambda_mano_joints * F.mse_loss(pred_mano_results['joints3d'], gt_mano_results['joints3d'])
loss['mano_pose'] = cfg.lambda_mano_pose * F.mse_loss(pred_mano_results['mano_pose'], gt_mano_results['mano_pose'])
loss['mano_shape'] = cfg.lambda_mano_shape * F.mse_loss(pred_mano_results['mano_shape'], gt_mano_results['mano_shape'])
loss['joints_img'] = cfg.lambda_joints_img * F.mse_loss(preds_joints_img[0], targets['joints_img'])
return loss
else:
# test output
out = {}
out['joints_coord_cam'] = pred_mano_results['joints3d']
out['mesh_coord_cam'] = pred_mano_results['verts3d']
return Transformer
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import repeat
class Transformer(nn.Module):
def __init__(self, inp_res=32, dim=256, depth=2, num_heads=4, mlp_ratio=4., injection=True):
super().__init__()
self.injection=injection
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(Block(dim=dim, num_heads=num_heads, mlp_ratio=mlp_ratio, injection=injection))
if self.injection:
self.conv1 = nn.Sequential(
nn.Conv2d(dim*2, dim, 3, padding=1),
nn.ReLU(),
nn.Conv2d(dim, dim, 3, padding=1),
)
self.conv2 = nn.Sequential(
nn.Conv2d(dim*2, dim, 1, padding=0),
)
def forward(self, query, key):
output = query
for i, layer in enumerate(self.layers):
output = layer(query=output, key=key)
if self.injection:
output = torch.cat([key, output], dim=1)
output = self.conv1(output) + self.conv2(output)
return output
class Mlp(nn.Module):
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
self._init_weights()
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
def _init_weights(self):
nn.init.xavier_uniform_(self.fc1.weight)
nn.init.xavier_uniform_(self.fc2.weight)
nn.init.normal_(self.fc1.bias, std=1e-6)
nn.init.normal_(self.fc2.bias, std=1e-6)
class Attention(nn.Module):
def __init__(self, dim, num_heads=1):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.sigmoid = nn.Sigmoid()
def forward(self, query, key, value, query2, key2, use_sigmoid):
B, N, C = query.shape
query = query.reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
key = key.reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
value = value.reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
attn = torch.matmul(query, key.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
if use_sigmoid:
query2 = query2.reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
key2 = key2.reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
attn2 = torch.matmul(query2, key2.transpose(-2, -1)) * self.scale
attn2 = torch.sum(attn2, dim=-1)
attn2 = self.sigmoid(attn2)
attn = attn * attn2.unsqueeze(3)
x = torch.matmul(attn, value).transpose(1, 2).reshape(B, N, C)
return x
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., act_layer=nn.GELU, norm_layer=nn.LayerNorm, injection=True):
super().__init__()
self.injection = injection
self.channels = dim
self.encode_value = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, stride=1, padding=0)
self.encode_query = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, stride=1, padding=0)
self.encode_key = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, stride=1, padding=0)
if self.injection:
self.encode_query2 = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, stride=1, padding=0)
self.encode_key2 = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, stride=1, padding=0)
self.attn = Attention(dim, num_heads=num_heads)
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer)
self.q_embedding = nn.Parameter(torch.randn(1, 256, 32, 32))
self.k_embedding = nn.Parameter(torch.randn(1, 256, 32, 32))
def with_pos_embed(self, tensor, pos):
return tensor if pos is None else tensor + pos
def forward(self, query, key, query_embed=None, key_embed=None):
b, c, h, w = query.shape
query_embed = repeat(self.q_embedding, '() n c d -> b n c d', b = b)
key_embed = repeat(self.k_embedding, '() n c d -> b n c d', b = b)
q_embed = self.with_pos_embed(query, query_embed)
k_embed = self.with_pos_embed(key, key_embed)
v = self.encode_value(key).view(b, self.channels, -1)
v = v.permute(0, 2, 1)
q = self.encode_query(q_embed).view(b, self.channels, -1)
q = q.permute(0, 2, 1)
k = self.encode_key(k_embed).view(b, self.channels, -1)
k = k.permute(0, 2, 1)
query = query.view(b, self.channels, -1).permute(0, 2, 1)
if self.injection:
q2 = self.encode_query2(q_embed).view(b, self.channels, -1)
q2 = q2.permute(0, 2, 1)
k2 = self.encode_key2(k_embed).view(b, self.channels, -1)
k2 = k2.permute(0, 2, 1)
query = self.attn(query=q, key=k, value=v,query2 = q2, key2 = k2, use_sigmoid=True)
else:
q2 = None
k2 = None
query = query + self.attn(query=q, key=k, value=v, query2 = q2, key2 = k2, use_sigmoid=False)
query = query + self.mlp(self.norm2(query))
query = query.permute(0, 2, 1).contiguous().view(b, self.channels, h, w)
return query
HandOccNet: Occlusion-Robust 3D Hand Mesh Estimation Network
official repo: https://github.com/namepllet/HandOccNet
Introduction
HandOccNet
Backbone - Feature injecting transformer (FIT) - Self-Enhancing transformer (SET) - Regressor
Backbone
Feature injecting transformer (FIT)
softmax-based attention module - sigmoid-based attention module
Softmax-based attention module
Fs로부터 Fp와 가장 관련이 있는 정보를 추출함. Occlusion을 야기하는 특정 정보는 이 모듈을 통해서 Fs와 Fp의 강한 상관관계를 표현할 수 있음.
Sigmoid-based attention module
Figure (g)와 같이 상대적으로 맵핑된 값으로 인해 불필요하게 커져버린 correlation을 걸러주는 모듈이 필요하여 sigmoid-based attention module을 사용
Feature injection
query 정보를 residual connection과 함께 output 단계에서 사용한 기존의 transformer와는 달리, query 정보는 사라지고 값들의 정보는 비어진 곳으로 투입되기 때문에 injection라는 용어를 사용.
Self-Enhancing transformer (SET)
F_fit에 self-attention module을 태움. Figrue 5. 참조. 모든 특징은 전형적인 self-attention의 특징을 따라 모든 key와 query가 적어도 스스로와 한번은 연관성을 갖게 됌.
Regressor
Hand mesh를 추출하기 위해서 SET의 출력 feature을 MANO pose parameters(48)와 shape parameter(10) 의 mesh로 mapping. Single-block hourglass, 4개의 residual blocks 그리고 fully connected layer로 구성됌.
Experiments
FIT and SET
Architecture of FIT
고찰