Chapter 6: Recognition

• For large-scale similarity search and clustering, the GPU-enabled Faiss library (https://github.com/facebookresearch/faiss) developed by Johnson, Douze, and Jégou (2021) can scale to very large datasets.

• There are many open-source frameworks such as Classy Vision (https://classyvision.ai), TensorFlow Core (https://www.tensorflow.org/tutorials/images/classification), and MMClassification (https://openmmlab.com) for training and fine tuning image and video classification models. You can also upload your images to the Computer Vision Explorer (https://vision-explorer.allenai.org) to see how well popular computer vision models perform on them.

• Open-source frameworks for training and fine-tuning object detectors include the TensorFlow Object Detection API (https://github.com/tensorflow/models/tree/master/research/object_detection), PyTorch's Detectron2 (https://github.com/facebookresearch/detectron2), and OpenMMLab's MMDetection (https://openmmlab.com/codebase#MMDetection) (Chen, Wang et al. 2019).

C.2 Software

• Detectron2 also includes semantic and panoptic segmentation, which can also be found in TensorFlow Core (https://www.tensorflow.org/tutorials/images/segmentation) and many other libraries.

• OpenPose (Cao, Hidalgo et al. 2019) and DensePose (Güler, Neverova, and Kokkinos 2018) are two popular software packages for determining “stick figure” and dense pixel-labeled 3D pose from 2D images.

• Pointers to software for more specialized tasks such as face detection and recognition, pedestrian detection, video understanding, and vision and language can usually be found alongside the latest papers discussed in Chapter 6.