classes are not balanced; add weights to ease the classification process

para probar directamente con las RGB

https://amaarora.github.io/2020/06/29/FocalLoss.html
https://towardsdatascience.com/understanding-binary-cross-entropy-log-loss-a-visual-explanation-a3ac6025181a

additional readings: https://arxiv.org/pdf/1901.05555.pdf

https://discuss.pytorch.org/t/focal-loss-for-imbalanced-multi-class-classification-in-pytorch/61289/3

then use this code to insert

<img src="https://user-images.githubusercontent.com/3965650/93581270-3441cd80-f9a1-11ea-89d4-dd1776e04584.png" width="600">

there are some errors/discrepancy bewteen the OSM ground truth and the BEVS.

After a review process performed using check_osm_with_gt.py and the associated output stored in
sftp://[email protected]/media/augusto/500GBDISK/nn.based.intersection.classficator.data/check_osm

with the output that is like an image + a txt file; if the first image differs from the last image, there is an error;

• first: anchor,
• second: image that should be like the anchor, a positive,
• third: a negative
• fourth: the category of the anchor in the GT

['/home/malvaro/Documentos/DualBiSeNet/data_raw/2011_09_30_drive_0018_sync/OSM/0000000003.png'] is type: 6
['/home/malvaro/Documentos/DualBiSeNet/data_raw/2011_09_30_drive_0027_sync/OSM/0000000643.png'] is type: 6
['/home/malvaro/Documentos/DualBiSeNet/data_raw/2011_09_30_drive_0020_sync/OSM/0000000549.png'] is type: 4

Last IMG is the GT of the ANCHOR
anchor lat: 49.049534450773
anchor lon: 8.3965921942334
positive lat: 48.986718086668
positive lon: 8.3938004084313
negative lat: 49.05333930648
negative lon: 8.3969599629762

Alvaro ended up with the following issues

Secuencia 28
	frame 3378 
	frame 3379 
	frame 4028
	frame 4029
	
Secuencia 33
	frame 758 - 772
	frame 991 - 997
	
Secuencia 27 (10)
	frame 1055 - 1072
	frame 1220 - 1244
	frame 3850

Try to train the net from a simple problem to a more complex problem.

bajar complexidad de la red y ver si hace overfit todavia

it seems to me that the shape is more important than the color.. maybe we can give a try with

• no normalization
• in grayscale?

create a baseline not only with the resnext101/50 but also with something used in some benchmark of road segmentation

this is the second frame from 2011_09_30_drive_0033_sync/0000000121.099

it would be nice to have a dataloader accepting a parameter to load only the images in a certain range, reporting the values from the GT also to the json generated

{"label": 3, "random_Tx": -1.4850388501380514, "random_Ty": 1.8083226303537128, "random_Tz": 1.118288729815498, "random_Rx": 0.6539320951297398, "random_Ry": 9.02848900921241, "random_Rz": -1.941300663631473, "starting_points": 67458, "remaining_points": 67458, "mirrored": false, "debug": true, "path": "/home/malvaro/Documentos/DualBiSeNet/data_raw_bev", "bev_path_filename": "/home/malvaro/Documentos/DualBiSeNet/data_raw_bev/2011_09_30_drive_0033_sync/0000000121.099.png"}