ADEPT: advanced dense prediction

Semantic analysis of natural images at the pixel level

Project ADEPT studies methods for dense semantic analysis in large images of natural scenes. Our goal is to reduce obstacles towards exciting real-world applications such as autonomous cars, road safety inspection or automated warehouses. We are especially interested in the following problems:

Research topics

Construction of universal taxonomies Dense hybrid anomaly detection Detection of anomalous regions
bevandic22bmvc grcic22eccv grcic223cvprw
[bevandic23bmvc] [grcic22eccv] [grcic23cvprw]

Training with overlapping labels Dense semantic forecasting with F2MF Densely connected normalizing flows
bevandic22wacv saric21tnnls grcic21neurips
[bevandic22wacv] [saric21tnnls] [grcic21neurips]


We study convolutional models for dense prediction based on checkpointed DenseNet backbones, lightweight ladder-style upsampling [1] and pyramidal fusion [2]. We predict future semantic content in video based on feature-to-motion-and-feature forecasting [3]. We apply multi-domain loss expressed as negative log likelihood of aggregated probability [4]. We address open-set recognition through learning with noisy and artificial negative samples [5].

[1] ladder-DenseNets, T-ITS 2020.
[2] pyramidal fusion, CVPR 2019, PR 2021.
[3] F2MF forecasting, CVPR 2020.
[4] multi-domain training, NLL+ loss, Arxiv.
[5] open-set recognition with noisy negatives, GCPR 2019.

Time frame

Start date: 1st February 2021.

Duration: 48 months

The project has been fully funded by the Croatian science foundation