Description
ImageNet (Deng et al., Russakovsky et al.) is the canonical large-scale image classification benchmark and the de facto pre-training corpus for modern computer vision backbones. In the vault its role is exclusively as the pre-training source for VGG / GoogLeNet / ResNet / ViT backbones that are then re-trained or fine-tuned for crowd counting, group counting, and CSI-vision fusion tasks.
Modality / size
- Modality: still RGB images with class labels.
- Scale: ~1.28M training images across 1000 classes (ILSVRC 2012 split).
Used by (papers)
- Backbone pretraining for crowd counting (CTRN, Wei et al. 2020) and group counting (Singh AOE, GoogLeNet/VGGNet pretrained on ImageNet).
- Cited as a reference open dataset that CSI-based behaviour-recognition datasets are aspired to match in scale.
Notes
- Standard public benchmark; download via the ILSVRC release.
- This note exists only because thesis-relevant papers use it as a backbone-pretraining reference; not used directly for CSI sensing.