Course projects

Project proposals for the course Visual Recognition with Deep Learning (taught in Spanish). Each project starts from one of the lab's models and datasets.

1. Improve sketch-based encoders for retrieval, segmentation and generation

Starting from skDINOv3, we will evaluate modifications of the encoder regarding training data, loss functions and architecture. The new encoder will be evaluated on retrieval, segmentation and detection.

General objective: propose an improved sketch-photo encoder.

Specific objectives:

  1. Evaluate existing encoders: skDINO and SketchFusion.
  2. Propose improvements to skDINO.
  3. Evaluate the model on segmentation and detection.
  4. Evaluate the sketch encoder on sketch-guided image generation.

Dataset · Model

2. Improve image encoders for historical documents (pattern detection and segmentation)

We will build an improved multi-scale encoder for historical document processing, based on this paper.

General objective: propose an improved encoder for pattern detection.

Specific objectives:

  1. Evaluate the iDoc encoder on retrieval.
  2. Propose improvements to iDoc.
  3. Evaluate the model on pattern detection.
  4. Evaluate the encoder on pattern generation.

Dataset · Model

3. Extend skDINO to sketch-based pattern spotting in historical documents

In this project we will propose a new encoder that detects patterns in historical documents from sketches.

General objective: propose a sketch-guided pattern detection method.

Specific objectives:

  1. Evaluate the skDINO encoder on retrieval.
  2. Propose improvements to skDINO for document patterns.
  3. Evaluate pattern detection with sketches.
  4. Evaluate the encoder on sketch-guided pattern generation.

Dataset · Model

4. Improve the medical imaging encoder for fine-grained anatomical structures in CT

We will evaluate and propose a CT encoder specialized in fine-grained structures, based on one of our recent works. A possible application is the segmentation of small tumours.

General objective: propose an improved encoder for localizing fine-grained structures in CT.

Specific objectives:

  1. Evaluate existing encoders, including our CT encoder.
  2. Define improvements to the CT encoder.
  3. Evaluate the encoder on small-tumour segmentation.
  4. Evaluate the encoder on CT generation.

Dataset · Model

5. Improve video encoders for news understanding

We will evaluate and improve video encoders applied to news description, building on one of our recent works.

General objective: propose an improved video encoder.

Specific objectives:

  1. Evaluate existing encoders.
  2. Define improvements to the best video encoder.
  3. Evaluate the model on short news videos.
  4. Evaluate video generation models.

Dataset · Model