yeah, they are so amazing

I have also checked it seems to be correct

'id_0660586f3ed26178d08c6636.jpg',

'id_0b2ce8235c041126ca62d9c1.jpg',

'id_0c2a4b8393c3c0dc26fe18bb.jpg',

'id_15fc1f078865759ff45638e6.jpg',

'id_169f24d5f7cb7b21fe60df1e.jpg',

'id_1c96db4131c3fab81949a9fd.jpg',

'id_22e2598aa43a8cd0a7fdb5df.jpg',

'id_2510cb8b891d54dd1570ecc5.jpg',

'id_253e58c20f3dfbcef9db86c5.jpg',

For example, can you check these?

It seems to be another worms? None of them when I put count for has improve my LB

i don't think it's another worms :v

oh really? So your opinion is ground truths are not correct?

yes, since they said they annotated them automatically by tool.

Here is the full list:

'id_0660586f3ed26178d08c6636.jpg',

'id_0b2ce8235c041126ca62d9c1.jpg',

'id_0c2a4b8393c3c0dc26fe18bb.jpg',

'id_15fc1f078865759ff45638e6.jpg',

'id_169f24d5f7cb7b21fe60df1e.jpg',

'id_1c96db4131c3fab81949a9fd.jpg',

'id_22e2598aa43a8cd0a7fdb5df.jpg',

'id_2510cb8b891d54dd1570ecc5.jpg',

'id_253e58c20f3dfbcef9db86c5.jpg',

'id_259990c7e44466c1853ac2fe.jpg',

'id_284f01098fe5bd92d93cc883.jpg',

'id_291c3927b05f73a336bd791f.jpg',

'id_2e7277d3546e5959d4c4b5ca.jpg',

'id_33dc6bd2f9f3c54ad6c03e1f.jpg',

'id_35d66ed5826a36d6e75053c6.jpg',

'id_38aed360b54c1e388e713440.jpg',

'id_447bca3dd59e31efab6719dd.jpg',

'id_4e186cb9161481297072821f.jpg',

'id_553f205c03ad2267d1073295.jpg',

'id_57701dfba47796b86b81e667.jpg',

'id_57eefc079173a65293e00c49.jpg',

'id_6055a3f37d25ac3a72d48b70.jpg',

'id_616bac757f4121ec2c319ff1.jpg',

'id_6737603b610159ab05ec356a.jpg',

'id_69378b6d40a23b51e72ed293.jpg',

'id_70986ad58af6664bf6fd5c6f.jpg',

'id_722d421d27db409c84cd90f1.jpg',

'id_75ec767e2219455019c1479c.jpg',

'id_78c79994c0fc9a2b4ccde17a.jpg',

'id_7e51b6cd7c803c9be4fbd0ab.jpg',

'id_8018014190322c8b2aa2cc25.jpg',

'id_8568d9e1533d530d168c7a04.jpg',

'id_8d0c3546d3ec9335339fbe32.jpg',

'id_8ddb685e1e2e2541bef9e6b1.jpg',

'id_91c21353aad56c62e5663129.jpg',

'id_9782d37ca668b5e66d4c3cfd.jpg',

'id_9a86018c21369815ffca0b07.jpg',

'id_a79b37b318ba77014b645ae0.jpg',

'id_a9cdb42497f9fd81c8e14880.jpg',

'id_b11d7e28d21c8c0b4507d94b.jpg',

'id_b8e7b3fe9aecae663a3cb687.jpg',

'id_ba6406be65dd3658d5d15bb9.jpg',

'id_bfda5d27ab0701de4cd48a40.jpg',

'id_bff124ee45fc42692656fd48.jpg',

'id_cba3ee362a28041a9c37641f.jpg',

'id_cf25b62de3e859747098a445.jpg',

'id_d696413c7f76d077d69bb605.jpg',

'id_e062aea6d9c3388be5e93ff1.jpg',

'id_e12b4de3a39da61ecadc349c.jpg',

'id_e76a6ba0911184c4bd8dd6b4.jpg',

'id_e86db6164f04c87a1a9ba7ed.jpg',

'id_eecae5c7d33eb81d2f308f8f.jpg',

'id_c9b824b2e88a356a98c4d25f.jpg',

'id_f89153e9ea4ee941935daef0.jpg'

Some may not change your MAE because it's in private test, but most of them will increase the MAE. I have manually check them by hand, and with yolo I can detect them as well, not another worms

If so, would you trust your model or the ground truth?

since they said in here: Worm count discrepancy - Zindi

could you trust the ground truth as there are errors in there?

I think the boxes is correct. May be some error when they count the number of worms lead to different? I still suppose that your worms list above are not contains pbw

Improves score for me

Maybe after the competition you guys share how you handled this problem? I am struggling breaking out from the 2.5-3 mae range. I would appreciate that. @mithilsalunkhe @tinyswish

I will surely share it once I get the time

Nothing special in my case, I just was fooled with public leaderboard) I'm also interested to hear about the winning solutions

I would be intrested in your solution too , how did you manage such a good public lb score

It was just first simple regression model without using bboxes etc and it surprisingly give me 3rd place on public lb with first submission) So, this is fooled me)

What is your solution?

Mine relied on yolov5

some kaggle GM magic :o

We've used swin-large, trained for 40 epochs, simple MAE, imgsize 384

We've tried beit_large with resolution 512, but it didn't improve the CV, as for the split, we first log-binarized the per-worm target and then stratified on these bins

But the best sub was with yolov7, detector works very well on private lb

im using dino by detrex, however, i am still doubting the ground truth, minor error can lead to different private score

I will share my solution later, but I've done overfit public LB :)