Child action recognition with GNN

objectives

first detailed ST-GNN based implementations on child action recognition

we have done detailed experiments on how intra-class and inter-class variations in child actions affect the model performance

we have utilized transfer learning methods to improve the performance on child action recognition

we develop a experimental method to measure the relationship between model accuracy and pose estimation noise and show that improved pose estimation result in improved model performance

experiment rationales

CWBG full: this is the main protocol/method for "child action recognition"

CWBG-dissimilar and CWBG-similar: these 2 for studying how inter-class variation in CAR affect the model performace

CWBG shared: this is the 5 shared classes. this is used to compare how train adult/test child VS train child/test child VS TFL with adult/test child VS large-diverse source dataset train/child test works basically compares how much improvement TFL can result in

intra-class variations

age wise - 3/4/5

gender wise male/female

Xsub vs random

check how if xsub is below random

this is what we expect since random has some person-related information shared -we expect random to perform better so

LOOCV

used to see how model works on each individual

can sort who is the best and worst; may give some insights

LOOCV is a better representation of the model also

ST-GNN

we are trying to see if improved/SOTA models can result in improvement :if so that means there is still potential for improved models to perform better

also results alone can be considered as a comparison

TFL methods

FX

FT

Hybrid

propagation

all these methods are used to see how far we can improve the model performance with source datasets

click to edit

ST-GCN architecture tuning -

this was done to see how much improvement can be achieved by changing the capacity of the model in depth and width wise

Research questions

Charith sir's feedback

click to edit

only within the given age group

between groups ex:3,4 vs 5

click to edit

GNN title justification: need experimental results to show that other skeleton methods perform worse than GNN

need to compare st-gcn as it is vs improved stgcn with child data

can use to analyse how age and gender groups perform on average

have to show evidence for all the Research questions

dharshana sir's feedback

need HAR(adult) vs child comparison

may be use NTU-5 for that

pradeepa madam feedback

include novelty

click to edit