PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution

Benchmark Model Rank Results
image-classification-on-imagenetPeLK-B-101–Top 1 Accuracy: 85.8%
object-detection-on-coco-val2017PeLK-B (Cascade Mask R-CNN 3x)–box AP: 52.9Params (M): 147
object-detection-on-coco-val2017PeLK-B-101 (Cascade Mask R-CNN 3x)–box AP: 53.1Params (M): 147
object-detection-on-coco-val2017PeLK-S (Cascade Mask R-CNN 3x)–box AP: 52.2Params (M): 108
object-detection-on-coco-val2017PeLK-T (Cascade Mask R-CNN 3x)–box AP: 51.4Params (M): 86
semantic-segmentation-on-ade20kPeLK-B-101–Validation mIoU: 50.6