Description for Human3.6M dataset.

H36M.ipynb

Human3.6M.md

Reference

@article{ionescu2014human3,
  title={Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments},
  author={Ionescu, Catalin and Papava, Dragos and Olaru, Vlad and Sminchisescu, Cristian},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  volume={36},
  number={7},
  pages={1325--1339},
  year={2014},
  publisher={IEEE}
}

Setup

• Download all archive files from the site.
  • You have to register with an academy mail.
• Extract all archive files.
  • NOTE: All training/test data must be extracted in the same path.
  • "The setup for our experiments requires that all the archives are extracted in the same directory (we call this the experiment directory)" quoted from ${code archive}/README.data file.
• Download and extract all dependencies described at ${code archive}/README file.
  • Neil Lawrence's mocap toolbox - required for visualizing poses
  • Mark Schmidt's great optimization package - required by our KDE inference
• Update VideoUtils from v.1.2 to v.1.2.4.
  • The memory release bug was fixed.
  • The maximum number of instances is increased from 128 to 1024.

Documantation

H36M/

H36MCamera.m

• R, T: Camera transform parameters
  • R: Euler angle
  • T: World position
• f, c, k p: Camera calibration parameters
  • f: Focal length
  • c: Principal point
  • k: Radial distortion
  • p: Tangential distortion

H36MCamera(db, s, c)

• db: H36MDataBase instance
• s: Subject number (camera is monocular if s == 0)
• c: Camera number (1 <= c <= 4)

project(obj, X)

• obj: Camera instance
• X: 3D position list

---

H36MDataAccess.m

An abstract class for H36MFeatureDataAccess, H36MPoseDataAccess, H36MTOFDataAccess and H36MVideoDataAccess classes. In H36MBL/H36MComputeFeatures.m, the Exists property is used to check if the precomputed data exists and the Permanent property is used to save the precomputed data if necessary.

• Exists: If the precomputed data exists
• Permanent: Save the precomputed data into the file

getFrame(obj, fno)

• obj: DataAccess instance
• fno: Frame number

---

H36MDataBase.m

The preloaded data from metadata.xml file. This includes camera parameters, joint relations and etc.

---

H36MFeature.m

An abstract class for H36MPose2DPositionsFeature, H36MPose3DAnglesFeature, H36MPose3DPositionFeature and etc.

---

H36MFeatureDataAccess.m

Called by H36MPose(2DPosition|3DPosition|3DAngles)Feature.serializer(Sequence), it reads whole precomputed feature data specified by the feature and the sequence.

• Buffer: Precomputed data

getFrame(obj, fno)

• obj: DataAccess instance
• fno: Frame number

---

H36MMyBBMask.m

Similar to the factory design pattern, generates the bounding-box mask DataAccess instance.

serializer(obj, Sequence)

Generates the bounding-box mask DataAccess instance.

• obj: H36MMyBBMask instance
• Sequence: Sequence instance

%% Sample code
sequence = H36MSequence(4, 3, 2, 1, -1)
bb_factory = H36MMyBBMask();
bb_handle = bb_factory.serializer(sequence);
bb_mask = bb_handle.getFrame(100);

---

H36MMyBGMask.m

Similar to the factory design pattern, generates the background mask DataAccess instance.

serializer(obj, Sequence)

Similar to H36MMyBBMask.serializer(obj, Sequence).

---

Sample codes

Visualize ToF

addpaths;
clear;
close all;

tof = H36MTOFDataAccess(6, 13, 2).getFrame(800); % subject, action, sub-action
figure(1); imshow(tof/10);

Issue

Regex error during the matlab version check

Delete the following code.

• ${code archive}/external_utils/xml_io_tools/xml_read.m

%% Check Matlab Version
v = ver('MATLAB');
version = str2double(regexp(v.Version, '\d.\d','match','once'));
if (version<7.1)
  error('Your MATLAB version is too old. You need version 7.1 or newer.');
end