ADEPT: advanced dense prediction

Semantic analysis of natural images at the pixel level


Multi-domain semantic segmentation with overlapping labels

  • we enable training and evaluation over incompatible taxonomies
  • our method expresses each dataset class as a union of disjoint universal classes.
  • the resulting model won the semantic segmentation contest at Robust Vision Challenge at ECCV 2020 and set the new state of the art on WildDash 2.
bevandic22wacv [bevandic22wacv]

Dense Semantic Forecasting in Video by Joint Regression of Features and Feature Motion

  • we express dense semantic forecasting as a causal relationship between the past and the future
  • we complement convolutional features with their respective correlation coefficients across a small set of discrete displacements
  • our single-frame model does not use skip connections along the upsampling path; hence we are able to forecast condensed abstract features at R/32
  • we present experiments for three dense prediction tasks: semantic segmentation, instance segmentation and panoptic segmentation`
saric21tnnls [saric21tnnls]

Densely connected normalizing flows

  • we show how to apply dense connectivity to normalizing flows
  • the resulting inductive bias assumes that some useful features are easier to compute than the others
  • however, straight-forward skip-connections preclude bijectivity
  • hence, we concatenate the noise after conditioning on previous representations
grcic21neurips [grcic21neurips]

A baseline for semi-supervised learning of efficient semantic segmentation models

  • we show that one-way consistency with clean teacher outperforms other forms of consistency (e.g. clean student or two-way) both in terms of generalization performance and memory efficiency
  • we propose a competitive perturbation model as a composition of a geometric warp and photometric jittering
  • we observe that simple consistency scales better than Mean Teachers in presence of more labels or more unlabeled data
  • we experiment on efficient models due to their importance for real-time and low-power applications.
grubisic22arxiv [grubisic21mva]