Hi, thanks for your wonderful work. I have a simple question that has confused me for a long time. If the source domain only has two lanes but the target domain has four lanes. What is the number of lanes we should set for the model to implement classification? I mention you set 2 in your experiment. But this may cause the left 2 lanes can not to be trained. Could you help solve my confusion?

Hi, thanks for the giant work. But whether there are a detail description of UFLD-SGPCS implement. Maybe it's a paper separate from “CARLANE: A Lane Detection Benchmark for Unsupervised Domain Adaptation from Simulation to multiple Real-World Domains”， bacause i don't seem to see the technical details in the paper CARLANE Benchmark

Looking forward to your reply !

Dear author,

Thank you for your work, my consider about UE4 coordinate system. What is the standard coordinate system and why we need to transform to this?
(x,y,z) -> (y,-z,x). New coordinate system makes me confuse.
If ue4 coordinate system left handed coordinate system with z up, so we just flip y axis to become right handed coordinate. So what extractly standard coordinate system that carla mentioned to?

Thank you.

Hi, when I am testing a pretrained model, I get very different accuracy numbers in different runs. Let's say I am using UFLD-ADDA. Below is the config file I used. Even before I train, the validate and test accuracy numbers vary quite a lot between different runs (sometimes by 4-6%).

Also, what is the difference between using 'pretrained' vs. 'finetune' in a config file? If I just want to finetune a pretrained model, should I just specify that model in pretrained or also in finetune?

molane config file:

DATA

data_root = '../UFLD-DANN/MoLane/data/'
source_train = '../UFLD-DANN/MoLane/splits/source_train.txt'
target_train = '../UFLD-DANN/MoLane/splits/target_train.txt'
source_val = '../UFLD-DANN/MoLane/splits/source_val.txt'
target_val = '../UFLD-DANN/MoLane/splits/target_val.txt'
target_test = '../UFLD-DANN/MoLane/splits/target_test.txt'

TRAIN

epoch = 1
batch_size = 16
optimizer = 'Adam' # ['SGD','Adam']
learning_rate = 1e-6
learning_rate_disc = 1e-3
weight_decay = 2.5e-5
momentum = 0.9
slope = 0.2

scheduler = 'cos' # ['multi', 'cos']

steps = [50,75]

gamma = 0.1
warmup = 'linear'
warmup_iters = 100

NETWORK

backbone = '18'
griding_num = 100
num_lanes = 2
cls_num_per_lane = 56
use_aux = False
num_workers = 4

LOSS

sim_loss_w = 1.0
shp_loss_w = 0.0

EXP

note = ''

log_path = './ufld_adda_log/'

FINETUNE or RESUME MODEL PATH

pretrained = '../pretrained/ufld_so_resnet18_ep149.pth'
finetune = None
resume = None

TEST

test_model = None
test_work_dir = None

Run 1 output:
[Validate on source]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 250/250 [01:37<00:00, 2.56it/s, top1=0.879, top2=0.968, top3=0.978]
[Validate on target]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 125/125 [00:49<00:00, 2.52it/s, top1=0.752, top2=0.868, top3=0.891]
[Validate Disc]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 125/125 [00:49<00:00, 2.53it/s, acc=0.422, correct=1687, total=4000]
[Test]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 63/63 [00:25<00:00, 2.48it/s, top1=0.759, top2=0.873, top3=0.882]

Run 2 output (without any change from run 1):
[Validate on source]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 250/250 [01:38<00:00, 2.55it/s, top1=0.866, top2=0.959, top3=0.972]
[Validate on target]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 125/125 [00:49<00:00, 2.55it/s, top1=0.753, top2=0.854, top3=0.873]
[Validate Disc]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 125/125 [00:49<00:00, 2.53it/s, acc=0.527, correct=2107, total=4000]
[Test]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 63/63 [00:24<00:00, 2.53it/s, top1=0.790, top2=0.901, top3=0.911]

Hi,

I am trying to generate pseudo labels in UFLD-SGADA but I get the following error:
python generate_pseudo_labels.py configs/molane.py
Traceback (most recent call last):
File "generate_pseudo_labels.py", line 364, in
encoder_state_dict = {key: value for key, value in adda_state_dict["encoder"].items() if "model" in key or "aux" in key or "pool" in key}
KeyError: 'encoder'

The pretrained model I am using is UFLD-SO downloaded from here for molane (ufld_so_resnet18_ep149.pth): https://carlanebenchmark.github.io/

Thanks!

I would like to express my gratitude for your contribution to the unsupervised domain adaptation for lane detection.

I have a question regarding the MoLane dataset, specifically about why the false positives (FP) and false negatives (FN) have equal values.

Thank you for your attention, and I look forward to your response.