420 likes | 528 Views
This document outlines the ETISEO annotation rules, data structures, and the processes for creating ground truth data for video object tracking. It details the context, object definitions, event specifications, and how to handle occlusions and artifacts during annotation. Various data formats and tools, including Viper-GT, are explained. The document is essential for ensuring consistency and accuracy in video annotations, facilitating effective analysis and reference data generation.
E N D
ETISEO Annotation rules • Data structure • Annotation tool and format • Ground truth creation rules • Reference data generation
Data structure - context • Context • Id • Type (ex: zone) • Name (ex: exit_zone) • Polygon
Data structure - object • Physical object • Id • Start time • Type (ex: person, vehicle…) • Subtype (ex: car, van…) • Info2D • Info3D • Links • Speed • GT Nature • GT Status • GT Occluded (partially or completely) • GT Connected with • GT : specific to annotation purpose
Data structure - event • Event • Id • Name • Start and end time • Status • Involved objects (physical object of interest and contextual object)
ETISEO Annotation rules • Data structure • Annotation tool and format • Ground truth creation rules • Reference data generation
Annotation tool and format • Annotation delivered with video: • Database - Excel file • Context annotation - Viper-GT tool and format • Annotation delivered at the end of Etiseo: • Ground truth - Viper-GT tool and format • Reference data (computed from GT) – Specific tool and Viper-GT format
Annotation tool and format • Viper-GT format example • <?xml version="1.0" encoding="UTF-8"?> • <viper xmlns="http://lamp.cfar.umd.edu/viper#" xmlns:data="http://lamp.cfar.umd.edu/viperdata#"> • <config> <…> • </config> • <data> • <sourcefile filename="file:ETI-VS1-CR-4-C1.avi"> • <object framespan="1:100" id="4" name="physical_object"> • <attribute name="nature"> • <data:svalue value="physical_object_of_interest"/> • </attribute> • <attribute name="type"> • <data:svalue value="person"/> • </attribute> • <attribute name="info2D"> • <data:bbox framespan="1:1" height="182" width="86" x="138" y="59"/> • <data:bbox framespan="2:2" height="183" width="86" x="139" y="59"/> • <data:bbox framespan="3:3" height="185" width="86" x="141" y="60"/> • <data:bbox framespan="4:17" height="187" width="87" x="143" y="60"/> • <…> • <data:bbox framespan="40:100" height="212" width="10" x="3" y="10"/> • </attribute> • <…> • </sourcefile> • </data> • </viper>
ETISEO Annotation rules • Data structure • Annotation tool and format • Ground truth creation rules • Reference data generation
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Contextual object • Annotated during its period of movement
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Static physical object of interest • Any physical object of interest is annotated on the whole video clip (ex: a parked vehicle) • First frame and last frame of a video clip
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Occlusions – general rules • Apparition of an object: the BBox increase with object apparition including only its visible part, • All BBox are limited by the image boundary, • Partial and complete occlusion will be differentiate in the evaluation (attribute “occluded”).
Ground truth creation – Rules for objects • Static occlusions (by a contextual object) • Partial occlusion: full BBox is drawn including hidden part,
Ground truth creation – Rules for objects • Static occlusions (by a contextual object) • Complete occlusion - two cases: • all “escapes” of the contextual object are visible: full BBox is drawn all the time,
Ground truth creation – Rules for objects • all “escapes” of the contextual object are not visible: the object is no more annotated when completely occluded
Ground truth creation – Rules for objects • Dynamic occlusion (by another object) • Partial or complete occlusion: in both cases, the full BBox is drawn including hidden part.
Ground truth creation – Rules for objects • Split object due to occlusion • A unique BBox including all sub part is drawn.
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Artefacts • Only the part constituting the objects forms the BBox, artefacts (shadow, reflection…) are ignored for instant. • Annotated in video data set 2
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Sudden illumination changes (tagged frames) • Static object • Object in movement: position interpolation
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Held object • The holder and the held object are both annotated (if visible), • The carried object is connected to the holder. • Example: bag-person, person-bike, door-vehicle
Ground truth creation – Rules for objects • Specific case • Person getting in/out a vehicle: annotated only when visible
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Group - Crowd • Group: multiple bounding boxes (persons + group) • Crowd : unique Bbox when individual annotation is impossible.
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Crowded video clip • Only object involved in event of interest are annotated. No false detections evaluation
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for objects • Multi-view annotation
Ground truth creation rules • Rules for objects: • Contextual object • Static physical object of interest • Occlusion • Artefacts • Sudden illumination changes • Held object • Group - Crowd • Crowded video clip • Multi-view annotation • Rules for events
Ground truth creation – Rules for events • Event – data structure • Id • Name • Start and end time • Status • Involved objects: • Id of the physical object of interest in GT files • Id of the contextual object in context information files
Ground truth creation – Rules for events • Choice of event for each video clip in the event ontology => defined in the database xls file (provided with video)
Ground truth creation – Rules for events • Automatic annotation of events with quantitative constraints, example: • Related to kinematics properties: stopped, running… • Related to spatial properties: close_to, inside_zone… • Manual annotation of other events: holding, pushing, waiting, gets_in/out,…
Ground truth creation – Rules for events • Example
ETISEO Annotation rules • Data structure • Annotation tool and format • Ground truth creation rules • Reference data generation
Reference data generation • Filtering on object size • Object speed computation • Group generation • Multi-view and 3D reconstruction • Events with quantitative constraints generation
ETISEO Annotation rules • Documentation: • ETI-Data structure and output format • ETI-Video Annotation Rules
ETISEO Annotation rules • Questions …