Could you please provide the pre-trained weights or the complete set of training parameters?

[awei-97]

trafficstars

Thank you for sharing your excellent work and making the code publicly available. I've attempted to reproduce the results using the training script in train_eval_gref.sh on the gref dataset, but encountered a significant performance gap。

My Implementation Results:

{'max_cIoU': 11.202466999449069, 'max_mIoU': 14.423027205342876, 'max_prec': 35.284451207956536, 'max_recall': 18.64305783005501, 'avg_cIoU': 26.584337117906927, 'avg_mIoU': 28.087503999400724, 'avg_prec': 33.12658978884035, 'avg_recall': 69.14119790617328} [avg_mIoU: 28.088] on [val set] of [Gref dataset]!!

After thoroughly reviewing, I have strictly followed the dataset processing methodology you provided and maintained all default parameters (only modifying the batch_size), yet the performance remains unsatisfactory.

Could you kindly:

Share the complete set of hyperparameters used for optimal training? Consider releasing a well-performing pretrained model to help verify implementation details?

[kdwonn]

Hi, sorry for the late reply. Seems like it's fairly well reproduced? (gref miou reported is 28.87) There could be various factors affecting model performance, like package versions, hardware, and seeds used, so I wouldn't consider this a failed reproduction by any means. (By the way, reducing batch size would be harmful for final performance since the model is using contrastive learning.) Anyway, I will upload the model weights soon and let you know when they're ready.