Federico Landi

I am a Ph.D. Candidate at AimageLab at the University of Modena and Reggio Emilia, Italy.

My research interests lie at the intersection of Computer Vision, Deep Learning, and Robotics, in the fascinating topic of Embodied AI. I work under the supervision of Prof. Rita Cucchiara.

As part of my Master Thesis, I was a visiting student at University of Amsterdam (UVA) where I worked under the supervision of Prof. Cees Snoek.

Email  /  CV  /  Google Scholar  /  Github  /  LinkedIn

In the first part of my Ph.D. I tackled the recent task of Vision-and-Language Navigation. More recently, I developed a strong interest in Embodied exploration and navigation, as well as in Recurrent Neural Networks.

We devise an impact-based intrinsic reward to train embodied exploration agents in photorealistic indoor environments.

A simple heuristic improvement boosts LSTM performance on a variety of tasks. Our approach shows more stable training dynamics and faster convergence time when compared to vanilla LSTM and peephole LSTM.

The first fully-attentive approach to Vision-and-Language Navigation (VLN). We achieve state-of-art performance on low-level VLN, a setting which is rarely considered in the literature given its additional difficulty.

We detail how to deploy a navigation policy trained on the Habitat simulator on a LoCoBot. Additionally, we study the performance on five different PointGoal navigation episodes in a real-world, challenging setting.

We present a new dataset of upper-body clothes for virtual try-on with high-resolution images. We also propose a new model for virtual try-on that can generate high-quality images using a three-stage pipeline.

We propose a new image-based virtual try-on model that can generate high-quality images. We exploit learnable affine and thin-plate spline transformations, combined with a generative network, to create new images of a person wearing different clothes.

A novel setting for Embodied AI in which an agent needs to explore a previously unknown environment while describing what it sees. The proposed model employs a self-supervised exploration module with penalty, and a fully-attentive captioning model for explanation.

In Vision-and-Language Navigation, an agent needs to reach a target destination with the only guidance of a natural language instruction. We exploit dynamic convolutional filters to ground the lingual description into the visual observation in an elegant and efficient way.

We explore the impact of considering spatiotemporal tubes instead of whole-frame video segments for anomaly detection in video surveillance. We create UCFCrime2Local: the first dataset for anomaly detection with bounding box supervision in both its train and test set.

Journals:

• IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)

• IEEE Robotics and Automation Letters (RAL)

• ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM)

• Pattern Recognition Letters (PRL)

Conferences:

• ACM International Conference on Multimedia (ACMMM)

• IEEE International Conference on Robotics and Automation (ICRA)

• IEEE International Conference on Pattern Recognition (ICPR)

• Computer Architecture - Prof. Rita Cucchiara, Prof. Simone Calderara, 2020-2021

• Machine Learning and Deep Learning - IFOA, 2020

• Deep Learning - Nuova Didactica, 2020

• Advanced Course on Data Science and Machine Learning - ACDL 2020, Remote (certificate)

• International Computer Vision Summer School - ICVSS 2019, Scicli (RG), Italy (certificate)

• In 2019, I carried on the 3D acquisition of the museum in Galleria Estense, in Modena (see below). One year later, the virtual spaces created for research purpose allowed to offer free guided tours to schools and young students during the Covid-19 lockdown in Italy.



• I like practicing Shuai Jiao (摔跤) and lifting weights.

I like this website.