Path |
|
|
save_model |
Path to save the model |
./saved_models/ |
load_model |
Path to load a model to continue training on it |
empty or ./saved_models/model1.h5 e.g. |
Data |
|
|
data_set |
Whether nuscenes, camra or coco for the specific data sets |
nuscenes |
data_path |
Path to the data set |
~/data/nuscenes |
save_val_img_path |
Path so save evaluated validation images after every epoch |
True / False |
n_sweeps |
Number of radar time steps used |
1 - 26 |
radar_projection_height |
Height of projected radar lines in m |
0.01 -> points, 1000 -> "barcode", or meters in between |
noise_filter_perfect |
Perfect noise filter based on ground truth |
True / False |
radar_filter_dist |
Filter radar data from a certain distance in m |
100 |
scene_selection |
Define validtion and test set |
default or debug |
Tensorboard |
|
|
tensorboard |
True if tensorboard logs should be saved |
True/False |
logdir |
Path to save tensorboard log files |
./tb_logs/ |
Computing |
|
|
seed |
Random seed to perform training |
e.g. 0 |
gpu |
Integer that specifies the system's GPU to run the training on |
e.g. 0 |
gpu_mem_usage |
Proportion (between 0 and 1)of GPU memory that should be used |
e.g. 0.5 |
workers |
Number of threads for generating data during training and evaluation |
e.g. 4 |
Preprocessing |
|
|
normalize_radar |
True if radar data should be normalized |
True/False |
random_transform |
True for extended data augmentation with rotation, shear etc. |
True/False |
sample_selection |
True to exclude samples without any objects from training |
True/False |
only_radar_annotated |
Only keep bounding boxes that have according radar points |
1 for nuScenes method, 2 for points_in_box method |
noisy_image_method |
Generate noisy image in data generator |
poisson, gauss, s&p-perpixel, blurr |
noise_factor |
Degree of noise, highly depends on the method |
e.g. 0.2 or 1e-4 |
Hyperparameters |
|
|
learning_rate |
Learning rate for training |
e.g. 1e-4 |
batchsize |
Batch size used for training |
e.g. 1 |
epochs |
Number of epochs for training |
e.g. 50 |
weighted_map |
True if mAP should be calculated weighted |
True/False |
category_mapping |
Categories (classes) to be used or merged |
see default.cfg |
class_weights |
Class weights for imbalanced classes |
see default.cfg |
CRF-Net |
|
|
channels |
Input channels (RGB + Radar) according to nuscenes |
e.g. 0,1,2,18 (encoding below) |
image_height |
Height of the input image in pixels |
e.g. 360 |
image_width |
Width of the input image in pixels |
e.g. 640 |
dropout_radar |
Chance that a sample has no radar data during training |
0 - 1 e.g. 0.2 |
dropout_image |
Chance that a sample has no image data during training |
0 - 1 e.g. 0.2 |
network |
Feature extractor network |
e.g. vgg-max-fpn or resnet101 |
network width |
Width factor of neural network to adapt number of kernels |
e.g. 0.5 or 1.5 |
pooling |
Pooling in radar branch |
max, min or conv |
anchor params |
default or small for different anchor sizes |
default or small |
pretrain_basenet |
True if feature extractor should initialized with ImageNet weights |
True/False |
distance_detection |
True if distances should be predicted (by an extra loss function) |
True/False |
distance_alpha |
Weight factor for distance loss |
e.g. 10 |
class_specific_nms |
True if NMS should be specific to classes |
True/False |
score_thresh_train |
Score trehsold, from which detections count as positive |
e.g. 0.05 |