Append 0 to default label names for single digit labels.

Aperta
#191 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
3/5
Tempo stimato
1-2 giorni
Idoneità per principianti
38/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Ferma

Direzione di ricerca

Inizia dalla generazione delle etichette predefinite dello strumento di labeling delle immagini e riproduci il problema con circa 15 etichette, quindi ispeziona l'XML utilizzato da train_shape_predictor.py. Il lavoro è completato quando le etichette predefinite a una cifra vengono scritte come 00 fino a 09 e l'ordine dei landmark rimane corretto quando l'XML viene letto per l'addestramento.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

The default label names for the first 10 labels are (1-9). However, they should be (00, 01, ... 09). Otherwise, when dlib reads the XML files and reorders the labels in ascending order the labels get messed up. For instance, 13 labels from (0-12) are mapped like the following by dlib before training. The right column is the original label and order of the landmarks in the .XML files. The left column is how the dlib training code perceives as their order. For example, the original 10th point is mapped to the 2nd point, the original 11th to 3rd etc. This results due to the missing 0 in front of 0-9 labels.

image

I used 10 images as both training and testing dataset. I achieved zero MAE on both testings and training. However, the order of the landmarks was all messed up. As I was getting 0 MAE that means the problem was happening with the XML reader of the dlib training functions.

I manually added the 0s to the labels and then the training was correct.
I know I can just manually add the 0s when labelling images using the tool. However, when going for thousands of images, this is a very unnecessary waste of time. I might for now write an XML parser script that will append the zeros.

To Reproduce
Steps to reproduce the behavior:

  1. Label a few images with the default labels. Use around 15 labels. Use the same images for training and testing.
  2. Train a shape predictor using http://dlib.net/train_shape_predictor.py.html
  3. Get zero MAE on both training and testing as we are just testing an overfitted model.
  4. See that the correct key points were detected but the ordering of the landmarks are messed up.

Screenshots
Ground truth:
image

Predicted Landmarks:
image

But MAE was zero. So that means DLIB is reordering the landmarks after reading. And then using that sorted list as the new ground truth order of the landmarks.

Lingua principale
HTML
Stelle
1k
Fork
611
Metriche di merge delle PR
Nessuna PR unita negli ultimi 30g

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di NaturalIntelligence/imglab

Tutte le issue di NaturalIntelligence/imglab

Issue simili

Altre issue su Machine Learning

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.