Proposal-contrastive pretraining for object detection from fewer data - Direction de la recherche technologique Access content directly
Conference Papers Year : 2023

Proposal-contrastive pretraining for object detection from fewer data


The use of pretrained deep neural networks represents an attractive alternative to achieve strong results with few data available. When specialized in dense problems such as object detection, learning local rather than global information in images has proven to be much more efficient. However, for unsupervised pretraining, the popular contrastive learning requires a large batch size and, therefore, a lot of resources. To address this problem, we are interested in Transformer-based object detectors that have recently gained traction in the community with good performance by generating many diverse object proposals. In this work, we propose ProSeCo, a novel unsupervised end-to-end pretraining approach to leverage this property of Transformer-based detector. ProSeCo uses the large number of object proposals generated by the detector for contrastive learning, which allows the use of a smaller batch size, combined with object-level features to learn local information in the images. To improve the effectiveness of the contrastive loss, we introduce the localization information in the positive selection to take into account multiple overlapping object proposals. When reusing pretrained backbone, we advocate for consistency in learning local information between the backbone and the detection head. We show that our method outperforms state-of-the-art in unsupervised end-to-end pretraining for object detection on standard and novel benchmarks in learning with fewer data.
Fichier principal
Vignette du fichier
full_paper.pdf (1.36 Mo) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

cea-04041965 , version 1 (22-03-2023)


  • HAL Id : cea-04041965 , version 1


Quentin Bouniot, Romaric Audigier, Angélique Loesch, Amaury Habrard. Proposal-contrastive pretraining for object detection from fewer data. ICLR 2023 - The Eleventh International Conference on Learning Representations, May 2023, Kigali, Rwanda. ⟨cea-04041965⟩
157 View
182 Download


Gmail Facebook X LinkedIn More