Lite-HRNet

The Lite-HRNet(Lite-HighResolutionNetwork) is a light weight backbone for high-resolution dependent computer vision tasks like Semantic Segmentation and Pose Estimation. It is a lightweight variant of HRNet(HighResolutionalNetwork), in which the authors replaces 1*1 convolutional blocks in HRNet with their innovative Cross Channel Weighting(CCW) to reduce both the time and spatial complexity. In this case, we implement the Lite-HRNet using Mindspore for the task of human pose estimation.

Performance and Pretrained Model

On COCO2017:

Arch	Input Size	#Params	FLOPs	AP	AP50	AP75	AR	AR50	ckpt
Naive Lite-HRNet-18	256x192	0.7M	194.8M	0.608	0.873	0.683	0.645	0.886	Mindspore Download
Wider Naive Lite-HRNet-18	256x192	1.3M	311.1M	0.643	0.893	0.725	0.680	0.901	Mindspore Download
Lite-HRNet-18	256x192	1.1M	205.2M	0.624	0.884	0.703	0.663	0.895	Mindspore Download
Lite-HRNet-18	384x288	1.1M	461.6M	0.668	0.895	0.739	0.702	0.909	Mindspore Download
Lite-HRNet-30	256x192	1.8M	319.2M	0.655	0.895	0.737	0.694	0.908	Mindspore Download
Lite-HRNet-30	384x288	1.8M	717.8M	0.698	0.905	0.781	0.731	0.918	Mindspore Download

On MPII:

Arch	Input Size	#Params	FLOPs	Mean	[email protected]	ckpt
Naive Lite-HRNet-18	256x192	0.7M	194.8M	0.842	0.270	Mindspore Download
Wider Naive Lite-HRNet-18	256x256	1.3M	311.1M	0.859	0.288	Mindspore Download
Lite-HRNet-18	256x256	1.1M	205.2M	0.842	0.262	Mindspore Download
Lite-HRNet-30	256x256	1.8M	319.2M	0.861	0.280	Mindspore Download

Prepare Dataset

You have to download COCO2017 dataset and annotation in COCODataset and make sure that the your path is organized as following:

Lite-HRNet/
    ├── imgs
    ├── src
    ├── annotations
        ├──person_keypoints_train2017.json
        └──person_keypoints_train2017.json
    ├── train2017
    └── val2017

Requirements

Train

python src/train.py

Optional Args	Default Value	Explanation
--model_type	lite_18	The type of model
--target_res	256x192	Resolution of resized input images
--checkpoint_path	./ckpts	Where to save or load checkpoints
--train_batch	32	Training batch size
--save_checkpoint_steps	500	The interval of saving checkpoints
--load_ckpt	False	Load a checkpoint and continue training

Evaluate

python src/eval.py

Optional Args	Default Value	Explanation
--model_type	lite_18	The type of model
--target_res	256x192	Resolution of resized input images
--checkpoint_path	./ckpts	Where to load checkpoints
--output_path	./eval_result	Where to save the evaluate result json file

Infer

python src/eval.py

Optional Args	Default Value	Explanation
--model_type	lite_18	The type of model
--target_res	256x192	Resolution of resized input images
--checkpoint_path	./ckpts	Where to load checkpoints
--infer_data_root	./infer_data	Where to load the input data
--out_data_root	./out_data	Where to save the output data

Result