CondLaneNet: a Top-to-down Lane Detection Framework Based on Conditional Convolution
Paper：https://arxiv.org/pdf/2105.05003.pdf
code：GitHub - aliyun/conditional-lane-detection
论文解读：
一、摘要
这项工作作为车道线检测任务，比较新颖的是检测头head。并不同于常规的基于bbox进行目标检测，这项工作采用的是检测关键点构造mask，输出形式类似instance segmentation。
二、网络结构
- backbone采用的是普通的CNN，比如ResNet；
- neck采用的是TransformerFPN，实际上就是考虑到车道线比较长，需要全局注意力，因此就在基础FPN构造金字塔之前对backbone输出的feature进行了Transformer的self-attention操作
- head分为两部分：
- Proposal head用于检测车道线实例，并为每个实例生成动态的卷积核参数；
- Conditional shape head利用Proposal head步骤生成的动态卷积核参数和conditional卷积确定车道线的point set。然后根据这些point set进行连线得到最后的车道线结果。
代码解析：
代码基于mmdetection框架(v2.0.0)开发。在config/condlanenet/里可以看到有三个文件夹，分别对应作者在三个数据集CurveLanes、CULane、TuSimple上的配置。它们之间最大的区别在于针对CurveLanes设计了RIM。下面我重点分析一下它们共同的一些模块：
backbone
采用的是resnet，根据模型的大小可能选择resnet18到resnet101不等
neck
这里采用的是TransConvFPN，在mmdet/models/necks/trans_fpn.py
跟FPN不同点主要在于多了个transformer操作。动机是觉得车道线比较细长，需要有self-attention这样non-local的结构。
也就是在resnet和FPN的中间多了一个transformer模块。
## TransConvFPN 不重要的代码部分已省略
def forward(self, src):
assert len(src) >= len(self.in_channels)
src = list(src)
if self.attention:
trans_feat = self.trans_head(src[self.trans_idx])
else:
trans_feat = src[self.trans_idx]
inputs = src[:-1]
inputs.append(trans_feat)
if len(inputs) > len(self.in_channels):
for _ in range(len(inputs) - len(self.in_channels)):
del inputs[0]
## 下面内容跟FPN一致
# build laterals
laterals = [
lateral_conv(inputs[i + self.start_level])
for i, lateral_conv in enumerate(self.lateral_convs)
]
## 省略
## 在TransConvFPN的__init__里
if self.attention:
self.trans_head = TransConvEncoderModule(**trans_cfg)
class TransConvEncoderModule(nn.Module):
def __init__(self, in_dim, attn_in_dims, attn_out_dims, strides, ratios, downscale=True, pos_shape=None):
super(TransConvEncoderModule, self).__init__()
if downscale:
stride = 2
else:
stride = 1
# self.first_conv = ConvModule(in_dim, 2*in_dim, kernel_size=3, stride=stride, padding=1)
# self.final_conv = ConvModule(attn_out_dims[-1], attn_out_dims[-1], kernel_size=3, stride=1, padding=1)
attn_layers = []
for dim1, dim2, stride, ratio in zip(attn_in_dims, attn_out_dims, strides, ratios):
attn_layers.append(AttentionLayer(dim1, dim2, ratio, stride))
if pos_shape is not None:
self.attn_layers = nn.ModuleList(attn_layers)
else:
self.attn_layers = nn.Sequential(*attn_layers)
self.pos_shape = pos_shape
self.pos_embeds = []
if pos_shape is not None:
for dim in attn_out_dims:
pos_embed = build_position_encoding(dim, pos_shape).cuda()
self.pos_embeds.append(pos_embed)
def forward(self, src):
# src = self.first_conv(src)
if self.pos_shape is None:
src = self.attn_layers(src)
else:
for layer, pos in zip(self.attn_layers, self.pos_embeds):
src = layer(src, pos.to(src.device))
# src = self.final_conv(src)
return src
class AttentionLayer(nn.Module):
""" Position attention module"""
def __init__(self, in_dim, out_dim, ratio=4, stride=1):
super(AttentionLayer, self).__init__()
self.chanel_in = in_dim
norm_cfg = dict(type='BN', requires_grad=True)
act_cfg = dict(type='ReLU')
self.pre_conv = ConvModule(
in_dim,
out_dim,
kernel_size=3,
stride=stride,
padding=1,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.query_conv = nn.Conv2d(
in_channels=out_dim, out_channels=out_dim // ratio, kernel_size=1)
self.key_conv = nn.Conv2d(
in_channels=out_dim, out_channels=out_dim // ratio, kernel_size=1)
self.value_conv = nn.Conv2d(
in_channels=out_dim, out_channels=out_dim, kernel_size=1)
self.final_conv = ConvModule(
out_dim,
out_dim,
kernel_size=3,
padding=1,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.softmax = nn.Softmax(dim=-1)
self.gamma = nn.Parameter(torch.zeros(1))
def forward(self, x, pos=None):
"""
inputs :
x : inpput feature maps( B X C X H X W)
returns :
out : attention value + input feature
attention: B X (HxW) X (HxW)
"""
x = self.pre_conv(x)
m_batchsize, _, height, width = x.size()
if pos is not None:
x += pos
proj_query = self.query_conv(x).view(m_batchsize, -1,
width * height).permute(0, 2, 1)
proj_key = self.key_conv(x).view(m_batchsize, -1, width * height)
energy = torch.bmm(proj_query, proj_key)
attention = self.softmax(energy)
attention = attention.permute(0, 2, 1)
proj_value = self.value_conv(x).view(m_batchsize, -1, width * height)
out = torch.bmm(proj_value, attention)
out = out.view(m_batchsize, -1, height, width)
proj_value = proj_value.view(m_batchsize, -1, height, width)
out_feat = self.gamma * out + x
out_feat = self.final_conv(out_feat)
return out_feathead
用的是CondLaneHead，在mmdet/models/dense_heads/condlanenet_head.py
需要重点分析，跟一般的检测任务差别很大：
首先这个CondLaneHead类的forward方法是直接调用了forward_test，因此要从model去看到neck输出后具体调用的是head的什么函数
# mmdet/models/detectors/condlanenet.py
def forward(self, img, img_metas=None, return_loss=True, **kwargs):
...
if img_metas is None:
return self.test_inference(img)
elif return_loss:
return self.forward_train(img, img_metas, **kwargs)
else:
return self.forward_test(img, img_metas, **kwargs)
def forward_train(self, img, img_metas, **kwargs):
...
if self.head:
outputs = self.bbox_head.forward_train(output, poses, num_ins)
...
def forward_test(self,
img,
img_metas,
benchmark=False,
hack_seeds=None,
**kwargs):
...
if self.head:
seeds, hm = self.bbox_head.forward_test(output, hack_seeds,
kwargs['thr'])
...所以实际上head的forward是没用到的，直接去看head的forward_train和forward_test就行
forward_train
# mmdet/models/dense_heads/condlanenet_head.py
def forward_train(self, inputs, pos, num_ins):
# x_list是backbone+neck输出后的multi level feature map
x_list = list(inputs)
# 这里根据hm_idx参数来取某个level 的feature map,用它去生成heat_map
# mask同理
f_hm = x_list[self.hm_idx]
f_mask = x_list[self.mask_idx]
m_batchsize = f_hm.size()[0]
# f_mask
z = self.ctnet_head(f_hm)
hm, params = z['hm'], z['params']
h_hm, w_hm = hm.size()[2:]
h_mask, w_mask = f_mask.size()[2:]
params = params.view(m_batchsize, self.num_classes, -1, h_hm, w_hm)
mask_branch = self.mask_branch(f_mask)
reg_branch = mask_branch
# reg_branch = self.reg_branch(f_mask)
params = params.permute(0, 1, 3, 4,
2).contiguous().view(-1, self.num_gen_params)
pos_tensor = torch.from_numpy(np.array(pos)).long().to(
params.device).unsqueeze(1)
pos_tensor = pos_tensor.expand(-1, self.num_gen_params)
mask_pos_tensor = pos_tensor[:, :self.num_mask_params]
reg_pos_tensor = pos_tensor[:, self.num_mask_params:]
if pos_tensor.size()[0] == 0:
masks = None
feat_range = None
else:
mask_params = params[:, :self.num_mask_params].gather(
0, mask_pos_tensor)
masks = self.mask_head(mask_branch, mask_params, num_ins)
if self.regression:
reg_params = params[:, self.num_mask_params:].gather(
0, reg_pos_tensor)
regs = self.reg_head(reg_branch, reg_params, num_ins)
else:
regs = masks
# regs = regs.view(sum(num_ins), 1, h_mask, w_mask)
feat_range = masks.permute(0, 1, 3,
2).view(sum(num_ins), w_mask, h_mask)
feat_range = self.mlp(feat_range)
return hm, regs, masks, feat_range, [mask_branch, reg_branch]forward_test
# mmdet/models/dense_heads/condlanenet_head.py
def forward_test(
self,
inputs,
hack_seeds=None,
hm_thr=0.3,
):
def parse_pos(seeds, batchsize, num_classes, h, w, device):
pos_list = [[p['coord'], p['id_class'] - 1] for p in seeds]
poses = []
for p in pos_list:
[c, r], label = p
pos = label * h * w + r * w + c
poses.append(pos)
poses = torch.from_numpy(np.array(
poses, np.long)).long().to(device).unsqueeze(1)
return poses
# with Timer("Elapsed time in stage1: %f"): # ignore
x_list = list(inputs)
f_hm = x_list[self.hm_idx]
f_mask = x_list[self.mask_idx]
m_batchsize = f_hm.size()[0]
f_deep = f_mask
m_batchsize = f_deep.size()[0]
# with Timer("Elapsed time in ctnet_head: %f"): # 0.3ms
z = self.ctnet_head(f_hm)
h_hm, w_hm = f_hm.size()[2:]
h_mask, w_mask = f_mask.size()[2:]
hm, params = z['hm'], z['params']
hm = torch.clamp(hm.sigmoid(), min=1e-4, max=1 - 1e-4)
params = params.view(m_batchsize, self.num_classes, -1, h_hm, w_hm)
# with Timer("Elapsed time in two branch: %f"): # 0.6ms
mask_branch = self.mask_branch(f_mask)
reg_branch = mask_branch
# reg_branch = self.reg_branch(f_mask)
params = params.permute(0, 1, 3, 4,
2).contiguous().view(-1, self.num_gen_params)
batch_size, num_classes, h, w = hm.size()
# with Timer("Elapsed time in ct decode: %f"): # 0.2ms
seeds = self.ctdet_decode(hm, thr=hm_thr)
if hack_seeds is not None:
seeds = hack_seeds
# with Timer("Elapsed time in stage2: %f"): # 0.08ms
pos_tensor = parse_pos(seeds, batch_size, num_classes, h, w, hm.device)
pos_tensor = pos_tensor.expand(-1, self.num_gen_params)
num_ins = [pos_tensor.size()[0]]
mask_pos_tensor = pos_tensor[:, :self.num_mask_params]
if self.regression:
reg_pos_tensor = pos_tensor[:, self.num_mask_params:]
# with Timer("Elapsed time in stage3: %f"): # 0.8ms
if pos_tensor.size()[0] == 0:
return [], hm
else:
mask_params = params[:, :self.num_mask_params].gather(
0, mask_pos_tensor)
# with Timer("Elapsed time in mask_head: %f"): #0.3ms
masks = self.mask_head(mask_branch, mask_params, num_ins)
if self.regression:
reg_params = params[:, self.num_mask_params:].gather(
0, reg_pos_tensor)
# with Timer("Elapsed time in reg_head: %f"): # 0.25ms
regs = self.reg_head(reg_branch, reg_params, num_ins)
else:
regs = masks
feat_range = masks.permute(0, 1, 3,
2).view(sum(num_ins), w_mask, h_mask)
feat_range = self.mlp(feat_range)
for i in range(len(seeds)):
seeds[i]['reg'] = regs[0, i:i + 1, :, :]
m = masks[0, i:i + 1, :, :]
seeds[i]['mask'] = m
seeds[i]['range'] = feat_range[i:i + 1]
return seeds, hm可以发现，这部分的操作跟论文中描述的差不多。
（等我具体有时间再慢慢弄来看，最近很忙）