Scribd
Upload
62 views

Retina Net

The document discusses using BiT models as backbones for the RetinaNet object detector and evaluating their performance on the COCO dataset. BiT models pretrained on larger datasets like ImageNet-21k and JFT-300M improve the Average Precision of RetinaNet by up to 1.5 and 2.1 points respectively compared to models pretrained on ImageNet only. RetinaNet is a one-stage object detector that uses Focal Loss to address class imbalance during training, allowing it to match the speed of previous one-stage detectors while surpassing the accuracy of two-stage detectors.

Uploaded by

pra3
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
62 views

Retina Net

The document discusses using BiT models as backbones for the RetinaNet object detector and evaluating their performance on the COCO dataset. BiT models pretrained on larger datasets like ImageNet-21k and JFT-300M improve the Average Precision of RetinaNet by up to 1.5 and 2.1 points respectively compared to models pretrained on ImageNet only. RetinaNet is a one-stage object detector that uses Focal Loss to address class imbalance during training, allowing it to match the speed of previous one-stage detectors while surpassing the accuracy of two-stage detectors.

Uploaded by

pra3
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 2

Big Transfer

For object detection, they use the COCO-2017 dataset and train a top-performing object detector,
RetinaNet (simple one-stage object detector), using pre-trained BiT models (BiT-S, BiT-M, BiT-S) as
backbones (share feature extraction layer). Due to memory constraints, they use the ResNet-101x3
architecture for all their BiT models. They fine-tune the detection models on the COCO-2017 train split
and do not use BiT-HyperRule but stick to the standard RetinaNet training protocol. It demonstrates that
BiT models outperform standard ImageNet pretrained models. They can see clear benefits of pre-
training on large data beyond ILSVRC-2012: pretraining on ImageNet-21k results in a 1.5 point
improvement in Average Precision (AP), while pretraining on JFT-300M further improves performance
by 0.6 points. (Refer to 201912.11370.pdf)

RetinaNet, named for its dense sampling of object locations in an input image. Its design features an
efficient in-network feature pyramid and use of anchor boxes. RetinaNet is a single, unified network
composed of a backbone network (in this case BiT models: BiT-S, BiT-M, BiT-S) and two task-specific
subnetworks. The backbone is responsible for computing a convolutional feature map over an entire
input image and is an off-the-self convolutional network. The first subnet performs convolutional object
classification on the backbone’s output; the second subnet performs convolutional bounding box
regression. While there are many possible choices for the details of these components, most design
parameters are not particularly sensitive to exact values as shown in the experiments. RetinaNet can be
trained with stochastic gradient descent (SGD). (Refer to 1708.02002.pdf)

Upstream Pre-Training they use BiT models as backbone: All of their BiT models use a vanilla ResNet-v2
architecture, except that they replace all Batch Normalization layers with Group Normalization and use
Weight Standardization in all convolutional layers. The first component is scale. It is well-known in deep
learning that larger networks perform better on their respective tasks. Further, it is recognized that
larger datasets require larger architectures to realize benefits, and vice versa. They study the
effectiveness of scale (during pre-training) in the context of transfer learning, including transfer to tasks
with very few datapoints. They investigate the interplay between computational budget (training time),
architecture size, and dataset size. For this, they train three BiT models on three large datasets: ILSVRC-
2012 which contains 1.3M images (BiT-S), ImageNet-21k which contains 14M images (BiT-M), and JFT
which contains 300M images (BiT-L). The second component is Group Normalization (GN)and Weight
Standardization (WS). Batch Normalization (BN)is used in most state-of-the-art vision models to stabilize
training. However, they find that BN is detrimental to Big Transfer for two reasons. First, when training
large models with small per-device batches, BN performs poorly or incurs inter-device synchronization
cost. Second, due to the requirement to update running statistics, BN is detrimental for transfer. GN,
when combined with WS, has been shown to improve performance on small-batch training for
ImageNet and COCO. Here, they show that the combination of GN and WS is useful for training with
large batch sizes and has a significant impact on transfer learning.

Transfer to Downstream Fine-Tuning Tasks they use RetinaNet protocol: Train all of their models for 30
epochs using a batch size of 256 with stochastic gradient descent, 0.08 initial learning rate, 0.9
momentum and 10−4 weight decay. They decrease the initial learning rate by a factor of 10 at epochs
number 16 and 22. They did try training for longer (60 epochs) and did not observe performance
improvements. The input image resolution is 1024 × 1024. During training they use a data augmentation
scheme as in [34, refer to 1405.0312.pdf]: random horizontal image flips and scale jittering. They set the
classification loss parameters α to 0.25 and γ to 2.0, see [33, refer to 1708.02002.pdf] for the
explanation of these parameters.

RetinaNet (one-stage detector) was designed to be more efficient than two-stage detectors. One-stage
detectors that are applied over a regular, dense sampling of possible object locations have the potential
to be faster and simpler, but previously have trailed the accuracy of two-stage detectors. It was
discovered that the extreme foreground-background class imbalance encountered during training of
dense detectors is the central cause. To address this class imbalance by reshaping the standard cross
entropy loss such that it down-weights the loss assigned to well-classified examples, Focal Loss method
was used which focuses training on a sparse set of hard examples and prevents the vast number of easy
negatives from overwhelming the detector during training. RetinaNet was used to evaluate the Focal
Loss method. Results show that when trained with the focal loss, RetinaNet is able to match the speed
of previous one-stage detectors while surpassing the accuracy of all existing state-of-the-art two-stage
detectors. (Refer to 1708.02002.pdf)

Applying the BiT models as backbone to RetinaNet showed clear benefits of one-stage detectors and the
benefits BiT models can be to object detectors by using standard training protocols.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy