about use column anchors for side lanes

[jing-turing]

I encountered some problems while testing in the experimental scenario.
this is my training config:

dataset= 'Tusimple'
data_root= 'project_kangaroo720p' # Need to be modified before running
epoch= 600
batch_size= 32
optimizer= 'SGD'
learning_rate= 0.05
weight_decay= 0.0001
momentum= 0.9
scheduler= 'multi'
steps= [50,75]
gamma= 0.1
warmup= 'linear'
warmup_iters= 100
backbone= '34'
griding_num= 100
use_aux= False
sim_loss_w= 0.0
shp_loss_w= 0.0
note= ''
log_path= ''
finetune= None
resume= None
test_model= ''
test_work_dir= ''
num_lanes= 4
var_loss_power= 2.0
auto_backup= True
num_row= 48
num_col= 41
train_width= 800
train_height= 320
num_cell_row= 100
num_cell_col= 100
mean_loss_w= 0.05
fc_norm= False
soft_loss= True
cls_loss_col_w= 1.0
cls_ext_col_w= 1.0
mean_loss_col_w= 0.05
eval_mode= 'normal'
crop_ratio = 0.8

Based on the Tusimple data format reference, the left image represents the result obtained from the trained model's inference, while the right image represents the ground truth.

In the bend, the inner side lane disappears.

But the straight sections remain normal.

Is this issue related to the "use column anchors for side lanes" mentioned in the paper? How can this problem be addressed?