In the TensorFlow Models Zoo, the object detection has a few popular single shot object detection models named "retinanet/resnet50_v1_fpn_ ..." or "Retinanet (SSD with Resnet 50 v1)". The paper usually linked to these works is here but the paper presents a different model, Detectron. So I understand SSD_ResNet50 FPN" uses the FPN feature extraction concept from that paper but are there any other more detailed documentation avaliable to understand how SSD integrates with FPN exactly? And what is the exact model architecture for this TF zoo Model?
Related
I am interested in building a yolo detector with trained on multiple datasets where each dataset has it own detection head. It is a multi-task learning approach. I am not sure how to convert the yolo detector architecture to support multiple head.
I came across the following projects, however I need your help to implement similar approach.
https://github.com/xingyizhou/UniDet
https://link.springer.com/chapter/10.1007/978-981-16-6963-7_27
This approach has some difficulties. First, in article you sent they use two-stage detection model with separate classification "branches". In the same time YOLO is one-stage detector and is fullyconvolutional, that means there are no fullyconnected layers, and class predictions (1d) are taking from the whole 3d-tensor (see the image).
You can take a look at YOLO9000 paper, the model was trained on detection and classification datasets at the same time - only loss function was changing.
TensorFlow Hub is a repository for pre-trained models. Model Garden (Model Zoo) also keeps SOTA models and provides facilities for downloading and leveraging its models like TfHub, and both of them are created by TensorFlow.
Why did Tensorflow make two concepts for a model repository?
When should we use TfHub for retrieving a well-known model, and when should we use Model Garden to download a model? What is the difference between them?
TF Hub provides trained models in SavedModel, TFLite, or TF.js format. These artifacts can be used for inference and some can be used in code for fine-tuning. TF Hub does not provide modeling library code to train your own models from scratch.
Model Garden is a modeling library for training BERT, image classification models, and more. Model Garden provides code for training your own models from scratch as well as some checkpoints to start from.
I'm starting to dive into tensor and object detection for a drone my friend and I are building. I keep seeing the word "model" thrown around and I'm sorry but I don't know what I should be picturing when I see the word "model" in terms of tensorflow and object detection.
Usually, in deep learning model is simply architecture of a neural network. It defines type of layers, number of nodes, connections, etc. Tensorflow uses static graph, which describes your model architecture in terms of nodes and operations. As a start you can use Keras API for defining your model.
https://keras.io/
Also read more about TF graph https://www.tensorflow.org/guide/graphs and take a look at tutorials https://www.tensorflow.org/tutorials
This tutorial discusses how to use objection detection API at tensorflow.
I am looking for the tutorial explaining how to add feature extractor such as mobilenetV2 to tensorflow's object detection framework.
Have you checked out the Tensorflow provided Model Zoo? :)
It includes various object detection models with various feature extractors such as MobileNet, Inception, ResNet etc.
Here below you will find a link to the Tensorflow Detection Model Zoo, where you can choose detection model architectures, Region-Based (R-CNN) or Single Shot Detector (SSD) models, and feature extractors.
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
You can download a frozen graph of the pre-trained models based on COCO, Kitti and Open-Images etc.
Tensorflow object detection API provides a number of pretrained object detection models to choose from. However, I would like to introduce modifications to the architecture of those models.
Particularly, I would like to make Faster RCNN into a more shallow network and use it to train my model. I want to gain in performance despite loss in accuracy. MobileNet is too inaccurate for my application.
Is it possible to achieve this without having to implement everything from scratch ?
Thank you.