Unsegmented Sports News Videos.
A collection of sports newscasts collected from youtube. The rapidly moving camera makes activity recognition challenging.
74 videos, Xvid codec, 240x320 pixels, 60-320 sec each.
10 classes: Swimming, American Football, Soccer, Tennis, Volley Ball, Track, Baseball, Golf, Basketball, Wrestling.
Average 3 sports/classes per video.
Tasks: Multi-label/Weakly supervised learning, Segmentation, Activity Detection, Classification, Localization.
A busy traffic intersection.
Video length: 1 hour (90000 frames)
Frame size: 360x288
Frame rate: 25 Hz
Compression codec: ffdshow mpeg-4
Tasks: Behaviour Profiling, Anomaly Detection, Background Subtraction.
Publications: ICCV'09, IJCV'11, PAMI'11.
Attribute Annotations Re-Identification: VIPER, PRID and GRID datasets.
Annotation of 15 attribute types for all 632 images in VIPeR pedestrian re-identification dataset.
Annotation of 21 attribute tpes for VIPER, PRID and GRID datasets.
Tasks: Re-identification, Soft-Biometrics, Attribute-Classification.
Publications: BMVC'12, ECCV ReID Workshop'12, Person Re-identification Book Chapter 2013.
Video feed from mobile drone.
Tasks: Person Reidentification, Reidentification in open world, Watchlist
Publications: ECCV'14 Workshop
Food-related Images, Tags and Metadata from Instagram Social Network
Tasks: Food Recognition and Health Analysis, Learning from Nosiy Tags, Social Network Analysis
Publications: ACM DH 2016
Learn a temporally correlated hierarchical topic model.
Learn a topic model for rare and subtle behaviours.
Exploratory data mining: Models for active learning and discovery / active learning in the presence of undiscovered classes. Contains the fusion method used in our PAKDD'11 and TKDE'11 papers as well as the new method in our ECCV'12 paper.
Deep Neural Network for Sketch Recognition: Train and test code for the CNN in our BMVC'15 paper.
Notes: 1. The framework used is matconvnet, so to use the the provided code you should either download the rather large provided packaged datafile, or write your own code to package the original TU-Berlin sketch data for matconvnet. 2. The provided model files contain all the trained models from the paper, however currently the provided test script only replicates the result of the simple non-ensemble version. To replicate the full result with ensemble it is necessary to implement or obtain the Joint Bayesian model from Chen et al's ECCV'12 paper.
Code [3kb] Trained Models [612MB] Matconvnet packaged Sketch data. With Stroke Order. [8GB] Matconvnet packaged Sketch data. Without Stroke Order. [3GB]
Multi-task multi-domain learning and zero-shot domain adaptation from our ICLR'15 paper.
In this paper we showed how many classic MTL/MDL algorithms can generalised by a class of neural nets. This also leads to the new zero-shot domain adaptation setting.
We have a good implementation using theano/python, as well as an illustrative example in matlab using the neural net toolbox.