20.500.12556/RUP-23158 Efficient material selection for training occluded mmWave radar-based gesture recognisers Radar-based gesture recognition has emerged as a promising approach for unobtrusive interaction. Unlike camera-based systems, radar sensors can detect gestures through opaque materials, enabling seamless embedding into various everyday objects. However, it remains unclear how to train models efficiently for robust gesture recognition through diverse materials. To investigate this, we collected a dataset of 17,520 gesture recordings performed through 73 everyday materials. By comparing several material-sampling and data-augmentation strategies, we found that a small carefully selected representative subset of training materials was sufficient to match performance of a classifier trained on the full material dataset. Our results showed that the models trained on 14 quota-sampled materials achieved accuracies of 95.8% and 91.2%, comparable to training on all 73 materials (96.8% and 91.6%) and significantly better than training without material data (66.8% and 65.8%). Among the evaluated sampling approaches, Quota sampling also provided the best overall trade-off between performance and practicality. In contrast, classifiers trained on augmented data performed worse than those trained on actual material-specific data. Taken together, these findings indicate that, for the tested sensor, gesture set, and material collection, carefully selected real-material data offer a practical route to reducing material-specific data collection in radar-based gesture recognition while preserving generalisation. Code, models, and data are available in the public repository, with additional details provided in the supplementary materials: https://gitlab.com/hicuplab/seeing-through. Radarsko prepoznavanje gest se je uveljavilo kot obetaven pristop za nemotečo interakcijo. Za razliko od sistemov, ki temeljijo na kamerah, lahko radarski senzorji zaznavajo geste skozi neprozorne materiale, kar omogoča neopazno vgradnjo v različne vsakdanje predmete. Vendar še ni jasno, kako učinkovito trenirati modele za robustno prepoznavanje gest skozi raznolike materiale. Da bi to raziskali, smo zbrali podatkovno zbirko 17.520 posnetkov gest, izvedenih skozi 73 vsakdanjih materialov. S primerjavo več strategij vzorčenja materialov in bogatenja podatkov smo ugotovili, da je majhna, skrbno izbrana reprezentativna podmnožica učnih materialov zadostovala za doseganje primerljive uspešnosti kot klasifikator, naučen na celotni podatkovni zbirki materialov. Naši rezultati so pokazali, da so modeli, naučeni na 14 materialih, izbranih s kvotnim vzorčenjem, dosegli točnosti 95,8 % in 91,2 %, kar je primerljivo z učenjem na vseh 73 materialih, kjer sta bili točnosti 96,8 % in 91,6 %, ter bistveno bolje kot učenje brez podatkov o materialih, kjer sta bili točnosti 66,8 % in 65,8 %. Med ovrednotenimi pristopi vzorčenja je kvotno vzorčenje ponudilo tudi najboljše splošno razmerje med uspešnostjo in praktičnostjo. Nasprotno pa so klasifikatorji, naučeni na obogatenih podatkih, dosegli slabše rezultate kot tisti, naučeni na dejanskih podatkih za posamezne materiale. Skupaj te ugotovitve kažejo, da pri testiranem senzorju, naboru gest in zbirki materialov skrbno izbrani podatki, pridobljeni z dejanskimi materiali, ponujajo praktično pot za zmanjšanje zbiranja materialno specifičnih podatkov pri radarskem prepoznavanju gest, hkrati pa ohranjajo zmožnost posploševanja. Koda, modeli in podatki so na voljo v javnem repozitoriju, dodatne podrobnosti pa so podane v dopolnilnih gradivih: https://gitlab.com/hicuplab/seeing-through. deep learning training optimization for neural networks millimetre-wave radar gesture recognition material selection sampling synthetic noise globoko učenje milimetrski radar prepoznavanje gest izbira materialov optimizacija treniranja nevronskih mrež vzorčenje sintetični šum true false true Angleški jezik Slovenski jezik Članek v reviji 2026-06-19 09:24:32 2026-06-19 09:24:37 2026-06-20 03:04:50 0000-00-00 00:00:00 2026 0 0 str. 1-21 article no. [Vol.] 2026 0000-00-00 PostprintKoncna NiDoloceno apc 0000-00-00 0000-00-00 2026-06-13 004.93 2045-2322 10.1038/s41598-026-56018-2 281590275 RAZ_Attygalle_Nuwan_2026.pdf RAZ_Attygalle_Nuwan_2026.pdf 1 0135C381F519077757A71EDC5050472F 3619fa6fd2ac806f1a4280273b593d742776ac88db6226a5a39786dbe492aa65 0bf8fe00-6bb0-11f1-9e8d-005056ac49c0 https://repozitorij.upr.si/Dokument.php?lang=slv&id=33852 https://www.nature.com/articles/s41598-026-56018-2 1 c82b1c18-6baf-11f1-9e8d-005056ac49c0 https://repozitorij.upr.si/Dokument.php?lang=slv&id=33851 Fakulteta za matematiko, naravoslovje in informacijske tehnologije 0 0 0