Warren Jouanneau's thesis: "Clothing identification in images: from instance segmentation to weakly supervised multi-label classification"

For the textile and fashion industry, clothing images are essential from the design of a product to its sale. It is therefore necessary to facilitate their access and search among many candidate images.  This is mostly done through the use of keywords to index them. The approach adopted in this thesis is divided into two steps: localization of each garment, then finer characterization of the content.
A study of instance segmentation methods was thus carried out. After having shown that the classical approaches of validation of segmentation masks are not adapted to our use cases, a three level evaluation protocol was proposed: global, contour, content.
Finally, it is sometimes difficult to classify an image with a single label. The multi-label classification of images is then more adapted. However, the application of supervised learning methods to this type of problem requires dedicated data, whose collection and annotation are complex. To address this problem, we have developed a weakly supervised method, requiring only one known positive label per image. The training then relies on a new efficient strategy to estimate negative examples for each class.

Thesis defended in person and possibility to follow it in videoconference, here.

Amphi LaBRI