Taskology: Utilizing Task Relations at Scale

Lu, Yao; Pirk, Sören; Dlabal, Jan; Brohan, Anthony; Pasad, Ankita; Zhao, Chen; Casser, Vincent; Angelova, Anelia; Gordon, Aaron

doi:10.1109/cvpr46437.2021.00859

Cited by 18 publications

(23 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Cross-task Relations. A rich body of work [3,8,27,34,39,49,54,62,[68][69][70]78] study the relations between tasks in MTL. Most related to ours, [49] explore the relations between segmentation and depth and propose a better fusion strategy to fuse two tasks predictions for domain adaptation.…”

Section: Related Workmentioning

confidence: 99%

“…the predictions made for multiple tasks from the same image are not independent, and therefore, are expected to be 'consistent'. Similar to [68], Lu et al [39] propose to leverage the cross-task consistency between predictions of different tasks on unlabeled data in a mediator dataset when jointly learning multiple models for distributed. To regularize the crosstask consistency, Lu et al [39] design multiple consistency losses according to the consistency between adjacent frames in videos, relations between depth and surface normal, etc.…”

Section: Related Workmentioning

confidence: 99%

“…Similar to [68], Lu et al [39] propose to leverage the cross-task consistency between predictions of different tasks on unlabeled data in a mediator dataset when jointly learning multiple models for distributed. To regularize the crosstask consistency, Lu et al [39] design multiple consistency losses according to the consistency between adjacent frames in videos, relations between depth and surface normal, etc. In this paper, we also exploit the cross-task consistency in MTL, however, from partially annotated data where the mapping from one task label to another cannot be analytically derived or exactly learned.…”

Section: Related Workmentioning

confidence: 99%

“…In this paper, we also exploit the cross-task consistency in MTL, however, from partially annotated data where the mapping from one task label to another cannot be analytically derived or exactly learned. To this end, unlike [39,68], we learn a joint task-space for each task pair rather than measuring consistency in one's task space. Finally, our method learns cross-task in a more computationally efficient way than [39,68] by sharing parameters across different mappings and conditioning its output on the related task-pair.…”

Section: Related Workmentioning

confidence: 99%

“…An orthogonal information that has recently been used in MTL is cross-task relation [39,49,68] which aims at producing consistent predictions across task pairs. Unfortunately existing methods are not directly applicable for learning from partial supervision, as they require either each training image to be labelled with all the task labels [49,68] or cross-task relations that can be analytically derived [39]. In our setting, there are fewer training images available with ground-truth labels of each task pair and thus it is harder to learn the relationship.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Learning Multiple Dense Prediction Tasks from Partially Annotated Data

Li¹,

Liu²,

Bilen³

2021

Preprint

View full text Add to dashboard Cite

Despite the recent advances in multi-task learning of dense prediction problems, most methods rely on expensive labelled datasets. In this paper, we present a label efficient approach and look at jointly learning of multiple dense prediction tasks on partially annotated data, which we call multi-task partially-supervised learning. We propose a multi-task training procedure that successfully leverages task relations to supervise its multi-task learning when data is partially annotated. In particular, we learn to map each task pair to a joint pairwise task-space which enables sharing information between them in a computationally efficient way through another network conditioned on task pairs, and avoids learning trivial cross-task relations by retaining high-level information about the input image. We rigorously demonstrate that our proposed method effectively exploits the images with unlabelled tasks and outperforms existing semi-supervised learning approaches and related methods on three standard benchmarks.

show abstract