Table 1 Performance of artificial neural networks from the timm library in the context of classification based temporal action localisation.
From: Temporal action localisation in video data containing rabbit behavioural patterns
Model | Number of parameters, millions | Average duration of 1 epoch at the training stage, s | Average duration of analysis of 1 batch of video frames at the stage of results evaluation, s |
|---|---|---|---|
EfficientFormer-L1 | 12.3 | 61.1 | 1.103 |
RegNetY-16GF | 11.2 | 124.3 | 1.097 |
DeiT-III-medium | 38.8 | 137.9 | 1.169 |
TresNet-L v.2 | 46.2 | 170.2 | 1.158 |
ConvNeXt-B | 88.6 | 717.5 | 1.266 |
ViT-base | 86.6 | 262 | 1.269 |
BEiT-large v.2 | 304.4 | 840.1 | 1.445 |
ViT-large | 304.2 | 1118.8 | 1.512 |