On Boosting Semantic Street Scene Segmentation with Weak Supervision

Meletis, Panagiotis; Dubbelman, Gijs

doi:10.48550/arxiv.1903.03462

Cited by 3 publications

(4 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This was the largest model with 448,794,235 trainable parameters and 87,356,100 non-trainable parameters. The optimizer of choice was ADAM [8] without gradient clipping and the loss function used was sparse categorical cross entropy [9] for all of the three models.…”

Section: Model Architecturementioning

confidence: 99%

Do you “Go big or go home” with Neural Networks?

Dahal

2022

Preprint

View full text Add to dashboard Cite

This paper includes study of three different text generation model approaches to test if making noticeable inferior changes in data preparation but significantly increasing the size of the model would have continuous drastic changes on the fidelity of results. I used chicken recipes found online to train text generative models that resulted in three different chicken recipes that no one may be able to follow. The general theme that emerged as the result of this experiment was: the outcome of a model is dictated not only by how big it is but also how well the data was processed for it.

show abstract

Section: Model Architecturementioning

confidence: 99%

Do you “Go big or go home” with Neural Networks?

Dahal

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…The optimizer of choice was ADAM [8] without gradient clipping and the loss function used was sparse categorical cross entropy [9] for all of the three models…”

Section: Model Based On 23mentioning

confidence: 99%

Do you “Go big or go home” with Neural Networks?

Dahal

2022

Preprint

View full text Add to dashboard Cite

This paper includes study of three different text generation model approaches to test impacts on feasibility in fidelity of results while making inferior changes to data preprocessing and simultaneously increasing the models size. Chicken recipes found online were used to train text generative models for evaluation. The general theme that emerged as the result of this experiment was: the outcome of a model is dictated not only by how big it is but by how well the data was processed for it.

show abstract

“…In this work we experiment with Cityscapes Dense subset [12] as C, and Cityscapes Coarse subset and Open Images bounding boxes subset [13] as O. We employ the hierarchical convolutional networks of [14] that can be trained on multiple datasets with strong and weak labels, which require the labels z of dataset O.…”

Section: Problem Definitionmentioning

confidence: 99%

“…We use our published hierarchical convolutional network for training simultaneously on weak and strong supervision for semantic segmentation [2], [14]. The network consists of a conventional ResNet-50 feature extractor, that is modified to have semantic segmentation output with dilated convolutions and an upsampling module.…”

Section: B Convolutional Model For Training On Strong and Weak Superv...mentioning

confidence: 99%

Data Selection for training Semantic Segmentation CNNs with cross-dataset weak supervision

Meletis

Romijnders

Dubbelman

2019

2019 IEEE Intelligent Transportation Systems Conference (ITSC)

Self Cite

View full text Add to dashboard Cite

Training convolutional networks for semantic segmentation with strong (per-pixel) and weak (per-bounding-box) supervision requires a large amount of weakly labeled data. We propose two methods for selecting the most relevant data with weak supervision. The first method is designed for finding visually similar images without the need of labels and is based on modeling image representations with a Gaussian Mixture Model (GMM). As a byproduct of GMM modeling, we present useful insights on characterizing the data generating distribution. The second method aims at finding images with high object diversity and requires only the bounding box labels. Both methods are developed in the context of automated driving and experimentation is conducted on Cityscapes and Open Images datasets. We demonstrate performance gains by reducing the amount of employed weakly labeled images up to 100 times for Open Images and up to 20 times for Cityscapes.

show abstract

On Boosting Semantic Street Scene Segmentation with Weak Supervision

Cited by 3 publications

References 0 publications

Do you “Go big or go home” with Neural Networks?

Do you “Go big or go home” with Neural Networks?

Do you “Go big or go home” with Neural Networks?

Data Selection for training Semantic Segmentation CNNs with cross-dataset weak supervision

Contact Info

Product

Resources

About