Focal Loss and Preprocessing Images

[Testbild]

Hello,

I have two questions regarding loss and preprocessing. Currently I have the issue, that the model detects objects very well, but assigns wrong classes to them. The objects I want to detect are left and right handed, so flipping the images in preprocessing/augmentation would actually let the model learn false classes. Hence my questions:

• For the retinanet the cls_loss is calculated with the focal_loss from libs/models/losses/losses.py focal_loss() correct?
• How can I be sure to have turned off image augmentation? I found in the read_tfrecord.py the read_and_preprocess_single_img() function, but could not find where it is called? In my cfgs I do not have any parameter regarding augmentation I believe:

from __future__ import division, print_function, absolute_import

import numpy as np

from libs.configs._base_.models.retinanet_r50_fpn import *
from libs.configs._base_.datasets.dota_detection import *
from libs.configs._base_.schedules.schedule_1x import *
from dataloader.pretrained_weights.pretrain_zoo import PretrainModelZoo

# schedule
BATCH_SIZE = 1
GPU_GROUP = "0"
NUM_GPU = len(GPU_GROUP.strip().split(','))
SAVE_WEIGHTS_INTE = 10000 * 2
DECAY_STEP = np.array(DECAY_EPOCH, np.int32) * SAVE_WEIGHTS_INTE
MAX_ITERATION = SAVE_WEIGHTS_INTE * MAX_EPOCH
WARM_SETP = int(WARM_EPOCH * SAVE_WEIGHTS_INTE)

# dataset
DATASET_NAME = 'myClass'
CLASS_NUM = 32

# model
# backbone
pretrain_zoo = PretrainModelZoo()
PRETRAINED_CKPT = pretrain_zoo.pretrain_weight_path(NET_NAME, ROOT_PATH)
TRAINED_CKPT = os.path.join(ROOT_PATH, 'output/trained_weights')

# bbox head
ANGLE_RANGE = 180

# loss
CLS_WEIGHT = 1.0
REG_WEIGHT = 1.0 / 5.0
REG_LOSS_MODE = 0

VERSION = 'RetinaNet_myClass'

• Is there maybe some other place where image augmentation is done?

Thank you very much and best regards!

EDIT: I also think I do not know, what exactly the ANGLE_RANGE is for? Maybe you could explain this also?

[yangxue0827]

ANGLE_RANGE denotes angle definition:

• opencv definition: [-90,0), ANGLE_RANGE=90
• long edge definition: [-90, 90), ANGLE_RANGE=180

[Testbild]

Thank you!