Instance segmentation with pixelwise convolutional neural network embeddings
(2023)
Files
thesis_final.pdf
Open Access - Adobe PDF
- 45.14 MB
Details
- Authors
- Supervisors
- De Vleeschouwer, Christophe
- Abstract
- In this thesis, we explore the use of pixelwise outputs predicted by convolutional neural networks to solve the Instance Segmentation (IS) task, i.e., the detection and contouring of individual objects in an image. IS constitutes a particularly challenging problem when images comprise numerous and occluded instances. We identify and compare the components constituting methods from the two main paradigms in bottom-up IS methods: spatial and associative embeddings. The former succeed in handling scenes with numerous instances, the latter in handling occlusions. Based on our comparisons, we engineer a hybrid method capable of handling the most difficult cases of occlusions in crowded scenes. In the course of this manuscript, we discuss the design of task formulations for neural networks, synthetic datasets, data-augmentation strategies, loss functions, post-processing pipelines, and modular software engineering constructs, that all have the potential to be transposed to a wide number of applications and domains.
- Affiliations
Citations
Istasse, M. (2023). Instance segmentation with pixelwise convolutional neural network embeddings. https://hdl.handle.net/2078.5/26978