Comparing with ExtremeNet and CornerNet

[vinbhaskara]

Since COCO hourglass starts with pretrained ExtremeNet weights which was in turn pretrained on CornerNet weights, is it fair to compare CenterNet to those models? Because effectively, CenterNet is trained for a lot more epochs when you include the epochs that went into the pretraining.

Also, can you point me to the pretrained weights that you used for training the hourglass model on COCO? Thanks.

[vinbhaskara]

One more query: Were the pretrained ExtremeNet weights (which you used to initialize Hourglass for COCO) trained originally with channels in RGB order (like other pytorch pretrained models) or BGR? Thanks.

[rafikg]

@bsvineethiitg did you check also this L1loss

CenterNet/src/lib/models/losses.py

Line 184 in a5a0483

loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean')

.
Here they average on the all pixels (128^2) but in the paper, they average on the non-zeros pixels (which is the true raisoning)!

[xingyizhou]

Thanks for the question. This turns out to be the major reason for this work being rejected. CenterNet training from scratch (100epochs) gives 38.9/ 40.6/ 43.0 for N.A./ flip/ multi scale. It is not all better than ExtremeNet (100epochs) 35.8/ 40.2/ 43.7.
In another perspective, finetuning ExtremeNet from the ExtremeNet (100epochs) pretraining improves ExtremeNet to 36.9/ 41.1/ 44.6, it also does worse than CenterNet in the same setting 40.9 / 42.8/ 45.7.
The CenterNet in the paper (40.3 / 42.2 / 45.1) is finetuned from ExtremeNet (50epochs) + CornerNet (~200epochs). We later observed that directly finetuning from ExtremeNet (100epochs) is better than that (40.9 / 42.8/ 45.7).
The ExtremeNet (100epochs) is in our model zoo.
The models are all trained in BGR.

[vinbhaskara]

Thanks for the question. This turns out to be the major reason for this work being rejected. CenterNet training from scratch (100epochs) gives 38.9/ 40.6/ 43.0 for N.A./ flip/ multi scale. It is not all better than ExtremeNet (100epochs) 35.8/ 40.2/ 43.7.

By "scratch" you mean ImageNet or MPII weights for Hourglass, right?

[chenyuntc]

Centernet is my favourite detector. It's ridiculous to reject it for this reason.

[xingyizhou]

Thanks @chenyuntc ! This means a lot to us.

[xingyizhou]

@bsvineethiitg "scratch" means using random initialization.

[vinbhaskara]

Hi @xingyizhou . Quick clarification please.

So, the model for COCO-HG you reported on the paper (40.3 / 42.2 / 45.1), you said that was finetuned from ExtremeNet (50epochs) + CornerNet (~200epochs).

I want to reproduce that.

Which pretrained model should I use for this? Is it "ExtremeNet_500000.pth"? I'm confused because you referred to that as the 100 epoch ExtremeNet model in your above comment.

Is ExtremeNet_500000.pth in your Drive link the 100 epoch ExtremeNet model or (50 epoch ExtremeNet starting from CornerNet 200 epoch) model?

[xingyizhou]

You are correct. ExtremeNet_500000.pth is the 100epoch model. The 50-epoch model ExtremeNet_250000.pth is in the ExtremeNet repo.

[vinbhaskara]

If you have trained your model on the 100 epoch ExtremeNet using the following command (with the same LR and #gpus and batch sizes):

python main.py ctdet --exp_id coco_hg --arch hourglass --batch_size 24 --master_batch 4 --lr 2.5e-4 --load_model ../models/ExtremeNet_500000.pth --gpus 0,1,2,3,4 --num_epochs 50 --lr_step 40

could you please share it in the google drive please? Also, I'd be glad if you can confirm if you trained that exactly with above command.

[wxiaoning]

If you have trained your model on the 100 epoch ExtremeNet using the following command (with the same LR and #gpus and batch sizes):

python main.py ctdet --exp_id coco_hg --arch hourglass --batch_size 24 --master_batch 4 --lr 2.5e-4 --load_model ../models/ExtremeNet_500000.pth --gpus 0,1,2,3,4 --num_epochs 50 --lr_step 40

could you please share it in the google drive please? Also, I'd be glad if you can confirm if you trained that exactly with above command.

Hello bsvineethiitg, have you tried this command? And can you reproduce the result?