Perception Encoder: The best visual embeddings are not at the output of the network

Benchmark Model Rank Results
image-classification-on-objectnetPE-Core G#1Top-1 Accuracy: 88.2
zero-shot-transfer-image-classification-on-sunPE-Core G#1Accuracy: 78.5
object-detection-on-coco-val2017PE_spatial (DETA)#2box AP: 66.0Params (M): 1900
zero-shot-cross-modal-retrieval-on-coco-2014PE-Core G#2Image-to-text R@1: 75.4
zero-shot-transfer-image-classification-on-food-101PE-Core G#2Top 1 Accuracy: 96.9
zero-shot-transfer-image-classification-on-imagenet-aPE-Core G#2Accuracy (Private): 92.6
zero-shot-transfer-image-classification-on-imagenet-rPE-Core G#2Accuracy: 96.5
zero-shot-transfer-image-classification-on-imagenet-sketchPE-Core G#2Accuracy (Private): 76.5
zero-shot-video-retrieval-on-msr-vtt-video-to-textPE-Core G#2R@1: 49.9
frozen-feature-label-propagation-on-davis-2017-valPE-Spatial G#5J&F: 61.5
instance-segmentation-on-lvis-v1-0-valPE-Spatial G + Mask R-CNN#5mask AP: 49.3
zero-shot-video-retrieval-on-msr-vttPE-Core G#6text-to-video R@1: 51.2
linear-probe-semantic-segmentation-on-ade20kPE-Spatial G#7mIoU: 48.9
image-classification-on-imagenet-v2PE-Core G#15Top 1 Accuracy: 80.2
zero-shot-video-question-answer-on-egoschema-fullsetPE-Lang G#23Accuracy: 62.0
instance-segmentation-on-coco-minivalPE-Spatial G + Mask R-CNN#26mask AP: 50.3
video-question-answering-on-mvbenchPE-Lang G#36Avg.: 69.4