SQLNet

Official Implementation for SQLNet: Scale-Modulated Query and Localization Network for Few-Shot Class-Agnostic Counting

Introduction

We propose a novel localization-based CAC approach, termed Scale-modulated Query and Localization Network (SQLNet). It fully explores the scales of exemplars in both the query and localization stages and achieves effective counting by accurately locating each object and predicting its approximate size.

Qualitative Results

The three columns are ground truth, predicted target points and predicted target bounding boxes.

Evaluation on Datasets

Result on FSC-147 dataset & Checkpoint:

Val MAE	Val MSE	Test MAE	Test MSE	Checkpoint
12.40	42.30	12.49	80.8	Download

Result on CARPK dataset:

	Method	MAE	MSE
Pre-trained on FSC-147	GMN	32.92	39.88
	FamNet	28.84	44.47
	BMNet+	10.44	13.77
	SAFECount	16.66	24.08
	SQLNet(ours)	7.66	9.66
Fine-tuned on CARPK	GMN	7.48	9.90
	FamNet	18.19	33.66
	BMNet+	5.76	7.83
	SAFECount	5.33	7.04
	SQLNet(ours)	4.89	6.55

Test

python test.py --vis --data_path /path/to/dataset --model_path ./path/to/pretrain_weight.pth

Note: Backbone(Resnet50) pre-trained parameters are automatically loaded and only part of them will be used by our model.

Code

• Test code & Pre-trained model
• Training code