CLIP-EBC: CLIP Can Count Accurately through Enhanced Blockwise Classification

Benchmark Model Rank Results
crowd-counting-on-nwpu-crowd-valCLIP-EBC (ViT-L/14)#2MAE: 32.3RMSE: 79.7
crowd-counting-on-nwpu-crowd-valCLIP-EBC (ViT-B/16)#3MAE: 36.6RMSE: 81.7
crowd-counting-on-nwpu-crowd-valCLIP-EBC (ResNet50)#4MAE: 38.6RMSE: 90.3
crowd-counting-on-nwpu-crowd-valDMCount-EBC#5MAE: 39.6RMSE: 95.8
crowd-counting-on-nwpu-crowd-valCSRNet-EBC#6MAE: 42.9RMSE: 100.1
crowd-counting-on-shanghaitech-aCLIP-EBC (ViT-B/16)#3MAE: 52.5RMSE: 85.9
crowd-counting-on-shanghaitech-aCLIP-EBC (ResNet50)#5MAE: 54.0RMSE: 83.2
crowd-counting-on-shanghaitech-aDMCount-EBC#15MAE: 62.3RMSE: 98.9
crowd-counting-on-shanghaitech-aCSRNet-EBC#17MAE: 66.3RMSE: 105.0
crowd-counting-on-shanghaitech-bCLIP-EBC (ViT-L/14)#2MAE: 5.9RMSE: 9.2
crowd-counting-on-shanghaitech-bCLIP-EBC (ResNet50)#4MAE: 6.0RMSE: 10.1
crowd-counting-on-shanghaitech-bCLIP-EBC (ViT-B/16)#8MAE: 6.6RMSE: 10.5
crowd-counting-on-shanghaitech-bCSRNet-EBC#11MAE: 6.9RMSE: 11.3
crowd-counting-on-shanghaitech-bDMCount-EBC#12MAE: 7.0RMSE: 10.9
crowd-counting-on-ucf-qnrfDMCount-EBC (16, dynamic)#2MAE: 75.90RMSE: 130.48
crowd-counting-on-ucf-qnrfDMCount-EBC (32, dynamic)#3MAE: 76.06RMSE: 127.72
crowd-counting-on-ucf-qnrfDMCount-EBC#4MAE: 77.2RMSE: 130.4
crowd-counting-on-ucf-qnrfCSRNet-EBC#5MAE: 79.3RMSE: 135.8
crowd-counting-on-ucf-qnrfCLIP-EBC (ViT-B/16)#7MAE: 80.3RMSE: 139.3
crowd-counting-on-ucf-qnrfCLIP-EBC (ResNet50)#8MAE: 80.5RMSE: 136.6