ICME 2023 Grand Challenge: PAIR競賽成績公告 The PAIR competition ended successfully

The “ICME2023 Grand Challenge – Low-power Deep Learning Object Detection and Semantic Segmentation Multitask Model Compression Competition for Traffic Scene in Asian Countries,” jointly organized by our center, MediaTek, iVS Lab (Intelligent Vision System Lab), AI System Lab, and A19 Lab, successfully concluded on April 19th. The competition aimed primarily at exploring the object image semantic segmentation technology for unique Asian traffic vehicles and road conditions with limited embedded resources. After one and a half months of preliminary rounds and three weeks of selection, as well as determining whether the papers were accepted by the ICME conference, the final winners have been announced.

Results of the evaluation in the final round

We have in total 9 teams uploaded their design in the final round and only five of them, i.e. Polybahn, You only lowpower once, asdgg, ACVLab, and 1000n-Huang Fan Club 2.0, whose models can be compiled and running at the target platform. All these five teams are invited to submit papers to ICME2023 PAIR session. Considering the paper review result as well as the completeness of the uploaded models, the associated materials and documents, only four teams are qualified to be the candidates of the winners, including Polybahn, You only lowpower once, ACVLab, and 1000n-Huang Fan Club.

In order to provide a fair evaluation on the four teams in the final round, we selected a set of 2285 images for the field as the true private dataset for the bounding box accuracy evaluation, along with the original private dataset for the semantic segmentation accuracy evaluation. The evaluation results are shown below:

Team Nmae	Total Score	Segmentation Accuracy (mIOU)		Detection Accuracy (mAP)		Model Complexity (GFLOPs)		Model Size (MB)		Inference Speed (us)		Power (mW)
		Raw	Score	Raw	Score	Raw	Score	Raw	Score	Raw	Score	Raw	Score
Polybahn	46.3	0.519	1.0	0.277	0.0	62.39	9.1	9.49	12.5	17,372	12.5	2808	11.1
You Only Lowpower Once	85.2	0.612	25.0	0.421	25.0	75.86	7.2	33.33	8.5	27,117	7.0	2772	12.5
ACVLab	36.8	0.515	0.0	0.334	9.9	38.62	12.5	83.51	0.0	26,595	7.3	2912	7.1
1000n-Huang Fan Club 2.0	23.0	0.543	7.2	0.316	6.8	126.84	0.0	30.10	9.0	39,602	0.0	3098	0.0

 

The ranking of the winners is based on the Overall Total score. The Champion title was awarded to the team “You Only Lowpower Once,” the 1st Runner-up was the team polybahn, and the 2nd Runner-up was ACVLab. The polybahn team achieved the best optimization results on the Mediatek platform, including a model size of 9.49MB, an inference speed of 17,372 us, and near-minimum power consumption, thus earning them the Best INT8 award.

 

 

Winners

• Champion：You Only Lowpower Once (Institute of Computer Science and Engineering, National Yang Ming Chiao Tung University)
• 1^st Runner-up：Polybahn (Mr. Chen Hong Ming)
• 2^nd Runner-up：ACVLab (Institute of Data Science, National Cheng Kung University)

 

Special Award

• Best INT8 model development Award：Polybahn (Mr. Chen Hong Ming)