CV3DST at TU München | Flashcards & Summaries

Select your language

Suggested languages for you:
Log In Start studying!

Lernmaterialien für CV3DST an der TU München

Greife auf kostenlose Karteikarten, Zusammenfassungen, Übungsaufgaben und Altklausuren für deinen CV3DST Kurs an der TU München zu.

TESTE DEIN WISSEN
Histogram of Oriented Gradients (HOG) (Method)
Lösung anzeigen
TESTE DEIN WISSEN
- Average gradient image over Training samples
- Gradients provide shape infromation
Lösung ausblenden
TESTE DEIN WISSEN

Histogram of Oriented Gradients (HOG) (Classification Steps)

Lösung anzeigen
TESTE DEIN WISSEN

1. Choose training set of images that contain object of interest

2. Choose a set of images that do NOT contain that object

3. Extracrt HOG features on both sets

4. Train an SQM classifier -> Detect whether or not feature vec is obj or not

Lösung ausblenden
TESTE DEIN WISSEN

Deformable Part Model (General)

Lösung anzeigen
TESTE DEIN WISSEN

- Basend on HOG features

- But classifies several defferent body parts


Lösung ausblenden
TESTE DEIN WISSEN

Faster R-CNN (Losses)

Lösung anzeigen
TESTE DEIN WISSEN

1. RPN classification (object/non-object)

2. RPN regression (anchor -> proposal)

3. Fast R-CNN classification (type of object)

4. Fast R-CNN regression (proposal -> box)

Lösung ausblenden
TESTE DEIN WISSEN

YOLO (General)

Lösung anzeigen
TESTE DEIN WISSEN

- Sliding window principle (here grid)

- Place box in the beginning in the center of the grid

- Direct regression from image to box coordinates

- For each grid location we predict n boxes

- 1 class prediction (20) + Anchor relative box regression (4) + Object / Non-object 



Lösung ausblenden
TESTE DEIN WISSEN

Single Shot multibox Detector (SSD) (general)

Lösung anzeigen
TESTE DEIN WISSEN

- Predicts at different scales

Lösung ausblenden
TESTE DEIN WISSEN

Problems with one-stage detectors:

Lösung anzeigen
TESTE DEIN WISSEN

- Many locations need to be analyzed (100k) densely coveing the image -> Foreground-background imbalance

- Hard negative mining is useful

Lösung ausblenden
TESTE DEIN WISSEN

Probelms with two-stage detectors:

Lösung anzeigen
TESTE DEIN WISSEN

- Classification only work in "interesting" foreground regions (proposals) -> Background examples are already filtered out

- But class balance between foreground and background objects is manageable

- Calssifier can concentrate on analyzing propsals with richt information content

Lösung ausblenden
TESTE DEIN WISSEN

RetinNet (general)

Lösung anzeigen
TESTE DEIN WISSEN

- One-Stage detector

- Changes the loss function -> Hard examples are weighted higher (Focal loss)

- Relies on ResNet for Feature extraction

- 9 anchors per level, each one with a classification and regression target


Lösung ausblenden
TESTE DEIN WISSEN

CornerNet (general)

Lösung anzeigen
TESTE DEIN WISSEN

- Express bounding boxes with 2 points, the top-left and bottom-rigth corners 

1. Probability map for each corner type (Heatmap) ( For every corner we have a different Heatmap)

2. Uses Embeddings to detect corners -> If two enmeddings are similar, then we can match them

3. (Class prediction)

- Predict corners at a lower resolution and then regress on offset -> Network learns relative position in low-resolution space



Lösung ausblenden
TESTE DEIN WISSEN

CornerNet (Cons)

Lösung anzeigen
TESTE DEIN WISSEN

- Many incorrect bounding boxes -> To many False Positives

- It is probably hard to infer the class of the box if the network is focused on the boundaries

Lösung ausblenden
TESTE DEIN WISSEN
Viola-Jones Detector (Steps)
Lösung anzeigen
TESTE DEIN WISSEN
1. Select your Haar-kind features (Handcrafted)
2. Integral image for fast feature evaluation
- Which parts of the image have the highest cross-correlation with my feature
3. AdaBoost to find weak learner
- Learn the best set of weak learners
- The final classifier is the linear combination of weak learners
Lösung ausblenden
  • 525001 Karteikarten
  • 11351 Studierende
  • 499 Lernmaterialien

Beispielhafte Karteikarten für deinen CV3DST Kurs an der TU München - von Kommilitonen auf StudySmarter erstellt!

Q:
Histogram of Oriented Gradients (HOG) (Method)
A:
- Average gradient image over Training samples
- Gradients provide shape infromation
Q:

Histogram of Oriented Gradients (HOG) (Classification Steps)

A:

1. Choose training set of images that contain object of interest

2. Choose a set of images that do NOT contain that object

3. Extracrt HOG features on both sets

4. Train an SQM classifier -> Detect whether or not feature vec is obj or not

Q:

Deformable Part Model (General)

A:

- Basend on HOG features

- But classifies several defferent body parts


Q:

Faster R-CNN (Losses)

A:

1. RPN classification (object/non-object)

2. RPN regression (anchor -> proposal)

3. Fast R-CNN classification (type of object)

4. Fast R-CNN regression (proposal -> box)

Q:

YOLO (General)

A:

- Sliding window principle (here grid)

- Place box in the beginning in the center of the grid

- Direct regression from image to box coordinates

- For each grid location we predict n boxes

- 1 class prediction (20) + Anchor relative box regression (4) + Object / Non-object 



Mehr Karteikarten anzeigen
Q:

Single Shot multibox Detector (SSD) (general)

A:

- Predicts at different scales

Q:

Problems with one-stage detectors:

A:

- Many locations need to be analyzed (100k) densely coveing the image -> Foreground-background imbalance

- Hard negative mining is useful

Q:

Probelms with two-stage detectors:

A:

- Classification only work in "interesting" foreground regions (proposals) -> Background examples are already filtered out

- But class balance between foreground and background objects is manageable

- Calssifier can concentrate on analyzing propsals with richt information content

Q:

RetinNet (general)

A:

- One-Stage detector

- Changes the loss function -> Hard examples are weighted higher (Focal loss)

- Relies on ResNet for Feature extraction

- 9 anchors per level, each one with a classification and regression target


Q:

CornerNet (general)

A:

- Express bounding boxes with 2 points, the top-left and bottom-rigth corners 

1. Probability map for each corner type (Heatmap) ( For every corner we have a different Heatmap)

2. Uses Embeddings to detect corners -> If two enmeddings are similar, then we can match them

3. (Class prediction)

- Predict corners at a lower resolution and then regress on offset -> Network learns relative position in low-resolution space



Q:

CornerNet (Cons)

A:

- Many incorrect bounding boxes -> To many False Positives

- It is probably hard to infer the class of the box if the network is focused on the boundaries

Q:
Viola-Jones Detector (Steps)
A:
1. Select your Haar-kind features (Handcrafted)
2. Integral image for fast feature evaluation
- Which parts of the image have the highest cross-correlation with my feature
3. AdaBoost to find weak learner
- Learn the best set of weak learners
- The final classifier is the linear combination of weak learners
CV3DST

Erstelle und finde Lernmaterialien auf StudySmarter.

Greife kostenlos auf tausende geteilte Karteikarten, Zusammenfassungen, Altklausuren und mehr zu.

Jetzt loslegen

Das sind die beliebtesten StudySmarter Kurse für deinen Studiengang CV3DST an der TU München

Für deinen Studiengang CV3DST an der TU München gibt es bereits viele Kurse, die von deinen Kommilitonen auf StudySmarter erstellt wurden. Karteikarten, Zusammenfassungen, Altklausuren, Übungsaufgaben und mehr warten auf dich!

Das sind die beliebtesten CV3DST Kurse im gesamten StudySmarter Universum

CVS

Wayamba University of Sri Lanka

Zum Kurs
CVD

Hochschule Fresenius

Zum Kurs
CVC

University of Geneva

Zum Kurs
CVA

Springfield Technical Community College

Zum Kurs
CV

Universität Leipzig

Zum Kurs

Die all-in-one Lernapp für Studierende

Greife auf Millionen geteilter Lernmaterialien der StudySmarter Community zu
Kostenlos anmelden CV3DST
Erstelle Karteikarten und Zusammenfassungen mit den StudySmarter Tools
Kostenlos loslegen CV3DST