Titulo Estágio
Through the magnifying glass: improving small object detection
Áreas de especialidade
Sistemas Inteligentes
Engenharia de Software
Local do Estágio
Coimbra
Enquadramento
A quick look at an image or video is enough for a human to recognise and locate objects of interest. In the digital world, object detection is a computer vision technique for locating instances of objects in images or videos. With recent advancements in deep learning based computer vision models, object detection applications are easier to develop than ever before. Besides significant performance improvements, these techniques have also been leveraging massive image datasets to reduce the need for large datasets when developing custom solutions. In addition, with current approaches focussing on full end-to-end pipelines, performance has also improved significantly, enabling real-time use cases. However, object detection performance degrades for small objects, particularly on noisy and/or lower-resolution images.
At Critical Software we are developing an embedded computer vision system that relies on different object detection models to achieve its goals. In this project we have gained un understanding of the challenges in this area, and of the benefits of building custom models for specific tasks and particular environments. Our next goal is to address the problem of small object detection. As such, the goal of this internship is twofold. First, a survey of the state-of-the-art models specific for small object detection is required to have a complete overview of the current state of research in this topic, the data that is used, the approaches that are being implemented, and its advantages and drawbacks. Second, building on the knowledge gained from the survey, a detection model for small objects should be proposed that tries to improve on the current state-of-the-art results.
Objetivo
The internship has two different stages. The goal of the first stage is to survey the current literature on the small object detection issue. As a result of this stage it is expected a systematic review of the most recent approaches in computer vision to address this specific task, and a listing of the relevant data being used for model development and evaluation.
This stage is important to develop the knowledge on the most recent techniques that will enable the work of the following stage. The goal of the second stage is to propose a new method/model to address this task, comparing it with state-of-the-art models where possible. The outcome of this stage is the proposed model itself along with its description, and the analysis and comparison of the results with the literature results.
Plano de Trabalhos - Semestre 1
The internship has the following stages:
- Survey the current literature [result: target article listing, M1 to M2]
- Reading and Writing the State of the Art [result: state of the art, M1 to M5]
- Writing the internship proposal [result: internship proposal, M5 and M6]
Plano de Trabalhos - Semestre 2
The second semester comprises the following stages:
- Setting up the Development Environment [result: Development Environment, M6]
- Evaluate current models [result: SoA performance results, M7 to M8]
- New proposal [result: model and performance results, M8 to M10]
- Writing the internship report [result: internship report, M10 and M11]
Condições
Bolsa de dissertação
O presente projeto de dissertação prevê a atribuição de uma bolsa mensal de 450 euros (considerando tempo integral). Esta bolsa é paga mensalmente e pretende ser uma ajuda para as despesas de deslocação e alimentação do aluno durante este período.
De salientar que, em contexto de dissertação, o foco das mais-valias que o aluno poderá obter estão associadas à aquisição de conhecimentos científicos e desenvolvimento de competências inerentes à sua integração profissional.
É fornecido computador e posto de trabalho.
Confidencialidade
A informação transmitida pela CRITICAL Software no âmbito do projeto de Dissertação, incluindo documentos técnicos ou de gestão, diagramas, código ou outra informação relevante deve ser tratada com a máxima confidencialidade. O candidato a quem for atribuído o projeto de Dissertação deve assinar um acordo de obrigação de confidencialidade (NDA, Non Disclosure Agreement).
Observações
During the survey the student should collect information on the used datasets and the data itself if publicly available. Alternatively, there are several public object detection datasets available which can be used to train the model(s) developed during the internship. As the goal of this internship is to address small object detection a specific dataset may be built using the relevant categories from datasets such as Google OpenImages, CIFAR100, MIT CSAIL objects and scenes, ImageMonkey, Objects365, and COCO.
Orientador
Rui Miguel Lourenço Lopes
rui.lopes@criticalsoftware.com 📩