UT-Austin Computer Vision Group Software

Please find software links below, together with the associated papers.

Learning the latent "look" for fashion images

Style-coherent embedding learning with polylingual LDA (based on MALLET)

On-demand learning for deep image restoration

Encoder-decoder model for image restoration tasks (inpainting, pixel interpolation, deblurring, denoising) and on-demand training approach to deal with mix of corruption difficulty levels

Object-centric representation learning from unlabeled videos


Unsupervised visual representation learning from object proposals using unlabeled video data. Caffe/Python implementation.

Self-supervised image representation learning from motion

"Embodied" equivariant visual representation learning using motor signals accompanying unlabeled video data. Caffe implementation.

Fine-Grained Visual Comparisons with Local Learning


Learning linear SVM rankers through fine-grained local learning. Pre-trained Mahalanobis matrices included. Use in conjunction with the UT Zappos50K dataset. Matlab implementation.

Zero-shot learning with unreliable attributes

Snap point detection for egocentric video

An algorithm to detect snap points with a Web photo prior. MATLAB code.

Deformable spatial pyramid matching

Fast, dense (pixel-level) image correspondences.  Shown to improve both accuracy and speed of SIFT Flow and Patch Match algorithms.

Boundary-preserving dense local region extraction (BPLRs)

BPLR feature detector, plus some descriptors for the extracted BPLRs, including HOG, chordiogram, and color histogram.  Written in MATLAB and tested in Linux 32 and 64 bits.

Kernelized locality sensitive hashing (KLSH)

Kernelized hashing algorithm which allows sub-linear time search under an arbitrary kernel function.  MATLAB code.

Interactive image segmentation with sufficiently strong user input

An algorithm to predict the annotation modality that is sufficiently strong for accurate segmentation of a given image (Matlab and C++).

Active learning to rank: Diverse setwise low margin selection algorithm

An algorithm for active learning of a ranking function that selects useful sets of examples to be partially ordered by an annotator, plus MTurk interface code for the cascading partial order annotation collection

Tree of Metrics (ToM)

Code and test data for Tree of Metrics (ToM), which learns a local metric at each node of a semantic taxonomy with disjoint sparsity regularization.

Semantic Kernel Forest

Code and test data for semantic kernel forest, which learns a category-specific kernel from forest of metrics learned on multiple semantic taxonomies. 

Generating taxonomies

Codes to generate a taxonomy of target categories based on WordNet or attributes, used in Tree of Metrics or Semantic Kernel Forests

Shape sharing for object segmentation

Generating object segment hypotheses based on shape sharing approach.  Written in Matlab and C++ (Matlab wrapper provided).

Crowdsourcing annotation tool to mark intervals of interest in videos

Extends VATIC tool (Vondrick et al.) to allow annotation of time intervals where an action occurs.

Max-subgraph activity detection

Action detection in continuous video.  Fast branch-and-cut approach to identify the subvolume in a video that maximizes the linear classifier's score.
  Note that CPLEX student license is available here.

Active frame selection for video label propagation

A dynamic programming-based algorithm to actively choose frames for human annotation along with label propagation with minimum expected error. Written in Matlab and tested on 64 bit Linux

Efficient region search object detection

Fast branch-and-cut approach to identify the subregion in an image that maximizes the classifier's score, for family of additive classifiers.  Note that CPLEX student license is available here.

Key-segments video object segmentation

Unsupervised segmentation of main foreground objects in video.

Asymmetric image matching

Region-to-image matching method, written in Matlab and C++ (Matlab wrapper provided).

Budgeted batch active learning

Batch-mode active learning algorithm to select the set of examples that appear most informative to an SVM, such that their total associated annotation cost meets a given budget. 


The full pipeline for context-aware discovery with object-graphs.  Implemented in Matlab

Multi-task feature learning

Multi-level multi-class VOI active selection

Implementation of Value-of-Information active learning criterion for multi-level multi-class case, as proposed in the CVPR 2009 paper

MIL training of object categories and MIL Bag-reweighting code

Pyramid match kernel - John Lee's LIBPMK: a pyramid match toolkit

Relative importance of objects in images

Relative importance embedding learned from tagged images with KCCA.

Incremental SVM code