ImageNet Large Scale Visual Recognition Challenge

Russakovsky, Olga; Deng, Jia; Su, Hao; Krause, Jonathan; Satheesh, Sanjeev; Ma, Siying; Huang, Zhiheng; Karpathy, Andrej; Khosla, Aditya; Bernstein, Michael S.; Berg, Alexander C.; Li, Feifei

doi:10.48550/arxiv.1409.0575

Cited by 289 publications

(303 citation statements)

References 51 publications

Supporting

Mentioning

297

Contrasting

Unclassified

Order By: Relevance

“…For all our experiments, we adopt the "Base" ViT model architecture (Dosovitskiy et al, 2020) pretrained on Ima-geNet (Russakovsky et al, 2014). Unless differently specified, we use clips of size 8×224×224, with frames sampled at a rate of 1/16.…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Is Space-Time Attention All You Need for Video Understanding?

Bertasius,

Wang,

Torresani

2021

Preprint

390

318

View full text Add to dashboard Cite

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically different design compared to the prominent paradigm of 3D convolutional architectures for video, TimeSformer achieves stateof-the-art results on several major action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Furthermore, our model is faster to train and has higher test-time efficiency compared to competing architectures. Code and pretrained models will be made publicly available.

show abstract

Section: Methodsmentioning

confidence: 99%

“…We show these results in Figure 5, where we also compare our method with SlowFast R50 (Feichtenhofer et al, 2019b), and I3D R50 (Carreira & Zisserman, 2017) trained on the same subsets. We note that all 3 architectures are pretrained on ImageNet (Russakovsky et al, 2014).…”

Section: The Importance Of Pretraining and Dataset Scalementioning

confidence: 99%

Is Space-Time Attention All You Need for Video Understanding?

Bertasius,

Wang,

Torresani

2021

Preprint

390

318

View full text Add to dashboard Cite

show abstract

“…In this section, we assess the practicality of XOR query and the proposed algorithm using real data collected from Amazon Mechanical Turk. We designed a binary classification task using 600 images of dogs and cats sampled from ImageNet [25]. Each human intelligent task (HIT) was designed to include 20 degree-1 queries and 20 degree-4 XOR queries.…”

Section: B Real Datasetsmentioning

confidence: 99%

Binary Classification with XOR Queries: Fundamental Limits and An Efficient Algorithm

Kim

Chung

2020

Preprint

View full text Add to dashboard Cite

Crowdsourcing systems have emerged as an effective platform to label data and classify objects with relatively low cost by exploiting non-expert workers. To ensure reliable recovery of unknown labels with as few number of queries as possible, we consider an effective query type that asks "group attribute" of a chosen subset of objects. In particular, we consider the problem of classifying m binary labels with XOR queries that ask whether the number of objects having a given attribute in the chosen subset of size d is even or odd. The subset size d, which we call query degree, can be varying over queries. Since a worker needs to make more efforts to answer a query of a higher degree, we consider a noise model where the accuracy of worker's answer changes depending both on the worker reliability and query degree d. For this general model, we characterize the information-theoretic limit on the optimal number of queries to reliably recover m labels in terms of a given combination of degree-d queries and noise parameters. Further, we propose an efficient inference algorithm that achieves this limit even when the noise parameters are unknown.

show abstract

“…However, the reported approach has substantial shortcomings and cannot be used for fast conversion of pre-trained neural networks on mobile devices. First of all it has a requirement to train threshold on the full Im-ageNet dataset [19]. Besides, it has no examples demonstrating the accuracy of networks used as standards for mobile platforms.…”

Section: Quantization With Training / Fine-tuningmentioning

confidence: 99%

“…Adam optimizer [13] is used for training, and cosine annealing with the reset of optimizer parameters -for learning rate. Training is carried out on approximately 10% part of Im-ageNet dataset [19]. Testing is done on the validation set.…”

Section: Scaling the Weights For Mobilenet-v2 (With Relu6)mentioning

confidence: 99%

Fast Adjustable Threshold For Uniform Neural Network Quantization (Winning solution of LPIRC-II)

Goncharenko,

Denisov,

Alyamkin

et al. 2018

Preprint

View full text Add to dashboard Cite

The neural network quantization is highly desired procedure to perform before running neural networks on mobile devices. Quantization without fine-tuning leads to accuracy drop of the model, whereas commonly used training with quantization is done on the full set of the labeled data and therefore is both time-and resource-consuming. Real life applications require simplification and acceleration of quantization procedure that will maintain the accuracy of full-precision neural network, especially for modern mobile neural network architectures like Mobilenet-v1, MobileNet-v2 and MNAS.Here we present two methods to significantly optimize the training with quantization procedure. The first one is introducing the trained scale factors for discretization thresholds that are separate for each filter. The second one is based on mutual rescaling of consequent depthwise separable convolution and convolution layers. Using the proposed techniques, we quantize the modern mobile architectures of neural networks with the set of train data of only ∼ 10% of the total ImageNet 2012 sample. Such reduction of train dataset size and small number of trainable parameters allow to fine-tune the network for several hours while maintaining the high accuracy of quantized model (accuracy drop was less than 0.5%). Ready-for-use models and code are avail-able at: https://github.com/agoncharenko1992/FAT-fastadjustable-threshold.

show abstract

ImageNet Large Scale Visual Recognition Challenge

Cited by 289 publications

References 51 publications

Is Space-Time Attention All You Need for Video Understanding?

Is Space-Time Attention All You Need for Video Understanding?

Binary Classification with XOR Queries: Fundamental Limits and An Efficient Algorithm

Fast Adjustable Threshold For Uniform Neural Network Quantization (Winning solution of LPIRC-II)

Contact Info

Product

Resources

About