Bad performance on my own dataset

[Georgegal00]

Hello , thank you for making this work available. I am currently trying to train the network on my own dataset consisting of images collected in the carla simulator.I have collected about 15,000 total images for train, validation and test set.I am using this configuration file:
`# Training settings
seed: 0
exps_dir: 'experiments'
iter_log_interval: 100
iter_time_window: 100
model_save_interval: 10
backup:
model:
name: PolyRegression
parameters:
num_outputs: 28 # (4 lanes) * (1 conf + 2 (upper & lower) + 4 poly coeffs)
pretrained: true
backbone: 'efficientnet-b1'
pred_category: false
curriculum_steps: [0, 0, 0, 0]
loss_parameters:
conf_weight: 1
lower_weight: 1
upper_weight: 1
cls_weight: 0
poly_weight: 150 #300
batch_size: 16
epochs: 385
optimizer:
name: Adam
parameters:
lr: 1.0e-3
lr_scheduler:
name: CosineAnnealingLR
parameters:
T_max: 385

Testing settings

test_parameters:
conf_threshold: 0.5

Dataset settings

datasets:
train:
type: LaneDataset
parameters:
dataset: carla
split: train
img_size: [360, 640]
normalize: true
aug_chance: 0.9090909090909091 # 10/11
augmentations:
- name: Affine
parameters:
rotate: !!python/tuple [-10, 10]
- name: HorizontalFlip
parameters:
p: 0.5
- name: CropToFixedSize
parameters:
width: 1152
height: 648
root: "/data/dataset/GG"
test: &test
type: LaneDataset
parameters:
dataset: carla
split: test
max_lanes: 4
img_size: [360, 640]
root: "/data/dataset/GG"
normalize: true
augmentations: []

val: &val
type: LaneDataset
parameters:
dataset: carla
split: val
max_lanes: 4
img_size: [360, 640]
root: "/data/dataset/GG"
normalize: true
augmentations: []`
The annotations are as in tuismple format, so I am processing them with the same methods
. Running python -m lib.datasets.lane_dataset , the loaded lanes are correct, I attach some pictures.








I am experiencing problems in the convergence of the model, the loss of train slowly decreases , while the loss of validation fluctuates between 1.2 and 3 , while the accuracy fluctuates between 0.35 and 0.55, this is at epoch 50 of the train. Analyzing the various logs, it can be seen that the polynomial loss is the predominant one. I also visualized some predictions to see where it was actually going wrong and I noticed that in some images it is getting close, while for others it is off completely, I attach some pictures.



I would like to understand if it is a data problem, dataset too small? Or is some adjustment to the model needed? Or do I simply need to figure out what the best parameters are? Thank you very much for your attention, any suggestions or answers are welcome!