OPTIMOL: automatic Object Picture collecTion via Incremental MOdel Learning

Introduction

A well-built dataset gives a good starting point for advanced computer vision research. It plays a crucial role especially in comparison, evaluation and adaptation of the state-of-the-art computer vision algorithms. However, dataset collecting is very tedious and time-consuming. In this paper, a novel automatical dataset collecting and model learning approach is presented by using object recognition techniques in an incremental way. Our algorithm mimics the human learning process in such a way that, starting from a few training examples, the more confident data you incorporate in the training data, the more reliable decision can be made. We demonstrate our framework by automatically collecting much larger object category datasets for 22 randomly selected classes from the Caltech101 dataset plus a "penguin" class. Furthermore, we offer not only more images in a dataset, but also a robust object model and meaningful image annotation.

Contribution

1. We propose an iterative framework that simultaneously collects object category datasets and learns the object category models.
2. We have developed an incremental learning scheme. This memory-less learning scheme is capable of handling any arbitrarily large number of images, a vital property for large image datasets.
3. Our experiments show that our algorithm is capable of both learning highly effective object category models and collecting object category datasets far larger than that of Caltech 101 or LabelMe.

Illustration of the framework of the Object Picture collecTion via Incremental MOdel Learning (OPTIMOL) system. This framework works in an incremental way in the following loop: Once a model is learned, it can be used to do classification on the images from the web resource. If the image is classified as in this object category, it gets accepted and incorporated into the collected dataset. Otherwise, it will be discarded. The model will again be updated by the newly accepted images in current round. In this incremental way, the category model gets more and more robust. As a consequence, the collected dataset gets larger and larger with reliable images.

Accordion

Amanita

Bonsai

Euphonium

Face

Grand-piano

Inline-skate

Laptop

Menorah

Nautilus

Pagoda

Panda

Penguin

Pyramid

Revolver

Schooner

Scoccer-ball

Starfish

Stop-sign

Strawberry

Sunflower

Umbrella

Watch

...

Web-23 Raw dataset

Dataset citation: Li et al. CVPR 2007

Download dataset:

Part1(accordion, amanita, bonsai, euphonium), Part2(faceI), Part3(faceII)

Part4(grand-piano, inline-skate, menorah, nautilus, pagoda), Part5(laptop, panda)

Part6(penguin, pyramid, schooner), Part7(revolver, soccer-ball, stop-sign, strawberry)

Part8(watch), Part9(sunflower, umbrella, starfish)