💡 Solutions are Public

[vrv]

Dear Participants!

The solutions from the three leaderboard winners are now public .

Check them out below

	Lab_3I	Solution
	gaurav_singhal	Solution
	Camaro	Solution

Did you find some interesting approaches in their solutions? Please share them with the community below!

[gaurav_singhal]

Nice approach by Lab_3i. QueryInst with Swin.

[lapp09]

#1: QueryInst + Swin Large
#2: HTC + Swin Base
#3: QueryInst + Swin Base

Looking forward to next year when Swin V2 Giant is in the picture.

I’m not familiar with QueryInst. Is there some reason you knew it would preform well on this dataset when other models perform better on COCO? COCO minival Benchmark (Object Detection) | Papers With Code

Care to share any comments on how you tuned your HTC/QueryInst parameters?

[saidinesh_pola]

I thought that only queryinst and swin v2 could be possible but swin v2 is just above or below the training limit for colab pro which just took away my days of training without any results. I realized queryinst do well as a result of @Mykola_Lavreniuk’s insightful advice. I find it really surprising why no one used this sahi library which slices the images into small parts, which gave me +10 mAp in just +2 training epochs on my best model. This, in my opinion, cuts down on training time for this challenge by more than 50%. I’m hoping that this will be useful for subsequent iterations of the challenge.