search

SysCV / cascade-detr

[ICCV'23] Cascade-DETR: Delving into High-Quality Universal Object Detection
https://arxiv.org/abs/2307.11035
Apache License 2.0
93 stars 4 forks source link

Questions about EgoHands Dataset #3

Open JiahuiChen-GitHub opened 1 year ago

JiahuiChen-GitHub commented 1 year ago

Great job! I have some questions regarding the EgoHands dataset.

  1. In Table 1 of the paper, the EgoHands dataset is mentioned to contain 11k images, but the dataset I obtained only consists of 4.8k images.
  2. The Roboflow team made corrections to the EgoHands dataset and reorganized the division between the training and testing sets. Did you utilize the Roboflow‘s version? If not, how do you split the dataset?

Looking forward to your response！

ymq2017 commented 1 year ago

Hi, thanks for watching our work! We use the EgoHands dataset with 11k images from this Roboflow Link and the provided default train/val split. Roboflow has different versions for this dataset and we use the latest one. We will also organize the download links of all datasets when releasing the code.

JiahuiChen-GitHub commented 1 year ago

Thanks! Looking forward to you releasing the code.

JiahuiChen-GitHub commented 1 year ago

Hi! here's another question. We are trying to train a model on the UVO dataset. We found that there are 2 versions of UVO: v0.5 and v1.0. They differ in the numbers of annotations and categories. If we want to make a fair comparison with your cascade-detr, which version should we choose?

ymq2017 commented 1 year ago

We use UVO v0.5 with its default train/val split.

JiahuiChen-GitHub commented 1 year ago

Hi! What did you set the image resolution to when you trained the model? 800x1333 or using multi-scale training strategy?

ymq2017 commented 1 year ago

Hi, sorry for the late reply. We use the same augmentation as DN-DETR with a multi-scale training strategy. The code of cascade_dn_detr is released now and you can find the corresponding augmentation code here. Feel free to ask me if you have more questions.