Skip to content

Data Design for Caesar RoME IIF

Jump to bottom
Dehann Fourie edited this page Nov 8, 2021 · 44 revisions

Copyright 2021 NavAbility, All rights reserved.

Discussion here on which data will be supported and how, regarding performance and availability -- see reference discussions below.

VND, Small, Meta (softtype), big data

What/Why Level data

Level2 as data which cannot be unpacked by Neo4j Cypher queries. Level1 is only allowed to exist as JSONable thing when serialized. Level1 also contains derived products from some of Level2 elements to allow so-called foveation queries. Level1 also corresponds to summary data in a graph which is starved of Level2 and Level3 data. Level0 is a immutable skeleton containing even less than Level1; in contrast Level3 contains full blown bigData elements (but doesn't have to be all the bigData elements at once).

Tabulated Summary

Note reworking attempt with new tabulated approach, see original full details below.

Variable

Level0 Level1 Level2 Level3 e.g.
SkeletonDFGVariable x label,tags
summaryData DFGVariableSummary x
solverData Dict Var.NodeData{T} x x :default=>Var.ND(..), :parallel_1=>...
softtypename TBD String x MIGHT BE FIXED BY DFG 184
estimateDict Dict <:Abst.PPEst. x x x :default=>MeanMaxPPE(..), :parallel_1=>...
bigDataKeys Dict Vector{String} or set x x x ['Data1', 'Data2']
smallData Dict String x x
bigData Dict <: B.D.Element x
...

Draft Table attempt

Julia Type Graph DB L0 L1 L2 L3 e.g.
DFG SkeletonDFGVariable Struct x
DFGVariableSummary Struct x
DFGVariable Struct x
label Symbol Label (Indexed Property?) x x x x0,x1,l1
tags Vector{Symbol} Labels x x x VARIABLE, LANDMARK, POSE
timestamp DateTime Indexed Property x x Epoch timestamp
estimateDict Dict{Symbol, <: Abst.PPEst.} Satellite Node x x :default=>MeanMaxPPE(..), :parallel_1=>...
softtypename String Indexed? x Pose2, Point3
solverDataDict Dict{Symbol, Var.NodeData{T<:Softtype}} JSON BLOB x :default=>Var.ND(..), :parallel_1=>...
smallDataKeys? Vector{Symbol} or Set{Symbol} [String]? ? ['Data1', 'Data2']
- smallData Dict{String, String} JSON BLOB x "DataKey"=>"DataValue"
bigDataKeys Vector{Symbol} or Set{Symbol} [String]? x ['Data1', 'Data2']
- bigData Dict{Symbol, AbstractBigDataEntry} JSON BLOB x key=>FileBigDataEntry(key, "FileName")
- - bigData Payload - - x Data Blob
ready Int Int x
backendset Int Int x
BigData DB x

List of Important Data

Level 0 (aka skeleton): SkeletonDFG[Variable/Factor]

  • Label
  • Tags

Level 1: DFG[Variable/Factor]Summary

Should be quickly searchable over thousands of nodes, possibly returned in summary

  • Parametric Point Estimate (PPE) are convenience byproducts of data full marginal estimation:
    • Only needed to make sense of max/mean/modefit/... data -eg. max = [2,4,0.1]
    • softype describes format, :Pose2 means [x,y,theta]
    • Also please define the fields somehow (maybe in documentation), for example :max will be there and the name won't change to something else eg. :maximum DFG #158.
    • [If 'MAP_est is just the numbers]
  • Time - Timestamp of the pose [UTC Epoch] (see Caesar.jl#269)
    • get all poses between time = [t1,t2]
  • Label [x1,x2,l1,etc.] (see Caesar.jl#251)
    • get x1
    • get x35 to x63
    • get all poses that saw landmark l1
  • The 'tags' [RobotId, SessionId, Variable, Factor, Landmark, Pose, ...]
    • get all Landmarks
    • get all for session_xyz
    • maybe if it is not important enough to want it searchable it should not be a tag
  • 🔴 softtype name as ::Symbol,
  • Position, Orientation [MAP_est? Also other related such as Velocity etc.]
    • get all poses close to coordinate [5, 3]
  • Naturally, all combinations of the above, eg. all the Poses close to time=t1 that 'saw' landmark l1.

Level 2: DFG[Variable/Factor]

Still in graph, quickly accessible for use in calculations. Is this copied to the local (In Mememory) graphs for calculations as part of CSM.

  • VariableNodeData
    • Stores all values associated with inference
  • softtype and its meta data
    • Passed to user residual function for specialized computations there.
  • smalldata
    • Micro second timestamp, sensor timestamp used in calculations.

Level 3: <: AbstractBigDataEntry

🔴 Working so far: https://github.com/JuliaRobotics/DistributedFactorGraphs.jl/pull/131/files

Big Data, all data not stored in the graph itself. Various MIME or basic types (see Caesar.jl#274)

  • Image
  • Depth Image
  • Point Clouds
  • etc.

This looks like this in the code: https://github.com/JuliaRobotics/DistributedFactorGraphs.jl/blob/master/src/entities/DFGVariable.jl#L83-L96

Note the first-class citizens, and the abstraction of the solver data.

Potentially Separate Config Parameters

Currently undecided, and originally raised here (IncrementalInference.jl#143)

DFG usage

Note that for CloudGraphs DFG instances, we are going to make (userId, robotId, sessionId) required for a connection. So, user's will have to provide DB connection + scope of the instance (user, robot, session). This is needed to make sure the solver queries are efficient.

This will make multisession queries a special case, which must be addressed. Effectively, that means that we may need to make the functions more generic, but it won't change the data or affect efficiency, which is the current concern.

E.g. of creating a CloudGraphs-based connection and working in the graph:

# Create connection
cgDFG = CloudGraphsDFG("localhost", 7474, "neo4j", "test", #DB params
    "testUser", "testRobot", "testSession", # user/robot/session
    IncrementalInference.encodePackedType, # encoding/decoding
    IncrementalInference.getpackedtype,
    IncrementalInference.decodePackedType)

# Now can find :x0
exists(cgDFG, :x0)
# Or add a node
using RoME
addVariable!(cgDFG, :x0, Pose2)

Working Variable / Data Definition

See DistributedFactorGraphs.jl DFGVariable.jl

Legend

  • 🔴 new field ('redlines')

References

Clone this wiki locally