1 / 10

Using linking features in learning Non-parametric part models

Ammar Kamal Hattab ENGN2560 Final Project Proposal March 11, 2013. Using linking features in learning Non-parametric part models. Project Goal. Nk. Torso. Project Goal: implement Linking Features Algorithm to detect a set of parts of a deformable object. Examples:

prisca
Download Presentation

Using linking features in learning Non-parametric part models

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. AmmarKamalHattab ENGN2560 Final Project Proposal March 11, 2013 Using linking features in learning Non-parametric part models

  2. Project Goal Nk Torso • Project Goal: implement Linking Features Algorithm to detect a set of parts of a deformable object. • Examples: • detect human parts: head, torso,upper/lower limbs • detect facial landmarks: eyes, nose, mouth outlines, etc. • detect animal parts • … tll trl bll brl

  3. Linking Features Method • Step 1 Preprocessing : • Finding bounding box • Computing image features (SIFT descriptors on a regular grid) Preprocessing Input Image SIFTs

  4. Linking Features Method • Step 2 Individual Parts: • SIFTs are grouped to generate part candidates for each part. • Each part candidate features vote for three orientations and one length estimate • Involves Generalized Hough Transform (GHT) • Using Training Images to help computeApproximate Nearest Neighbors featuresefficiently Preprocessing Individual parts Input Image SIFTs part candidates

  5. Linking Features Method • Step 3: Pairwise Connectivity • Find ‘linking features’ for each pair part candidates using Training (to find nearest neighbors) • Find the correct choice of part candidates: • Score of a pair of part candidates is measured by the cumulative score of the linking features between them • This results in combined set of estimated connectivity parameters Preprocessing Individual parts Input Image Pairwise Connectivity connectivity parameters part candidates

  6. Linking Features Method • Step 4: Finding the most likely global parts configuration • The most likely joint configuration of parts is inferred from the model. • the most consistent subset of candidate parts is computed by a greedy approximate MAP inference over the estimated model Preprocessing Individual parts Input Image Pairwise Connectivity Global Parts Configuration

  7. Project Plan • Stage 1: Implement Preprocessing (and finding SIFTs) • Stage 2: Implement Training (to help compute Approximate Nearest Neighbors Features) • Stage 3: Implement Finding parts candidates (Involves Generalized Hough Transform (GHT) and clustering) • Stage 4: Finding Linking Features using Training Images • Stage 5: Finding the most likely global parts configuration using greedy approach Preprocessing Individual parts Input Image Pairwise Connectivity Global Parts Configuration Parts

  8. Results Example

  9. References • Leonid Karlinsky, Shimon Ullman: Using Linking Features in Learning Non-parametric Part Models. ECCV (3) 2012: 326-339 • Felzenszwalb, P., Huttenlocher, D.: Pictorial structures for object recognition. IJCV (2005)

  10. Training Phase • The training phase of the method receives a set of images with annotated parts (e.g. stick annotation in upper body experiments). • Enlarge them and find each part descriptor • Then building efficient data structures for similar descriptor search (Approximate Nearest Neighbors ANN) and memorizing a set of parameters for each feature, e.g. relative offset from different parts. • These data structures are later used in order to compute the model probabilities (features probabilities) for test images using Kernel Density Estimation (KDE) Training Images A test feature KDE Fi GHT=

More Related