search

bytedance / coconut_cvpr2024

Apache License 2.0
151 stars 6 forks source link

COCO panoptic and Sama-COCO masks #11

Open darkasevgen opened 6 months ago

darkasevgen commented 6 months ago

Hello! I have a question. You wrote the following: We adopt COCO panoptic and Sama-COCO masks (high-quality instance segmentation annotations) as our starting point. But as far as I understand, in your labeling pipeline, you used in step 1 kMaX-Deeplab and DETA models to get masks and bbox proposals respectively. Hence, the COCO and Sama-COCO masks are not used, am I right?

xdeng7 commented 6 months ago

Hello! I have a question. You wrote the following: We adopt COCO panoptic and Sama-COCO masks (high-quality instance segmentation annotations) as our starting point. But as far as I understand, in your labeling pipeline, you used in step 1 kMaX-Deeplab and DETA models to get masks and bbox proposals respectively. Hence, the COCO and Sama-COCO masks are not used, am I right?

to build the step-1 data engine, we did not use sama-coco, but used it for the step2 semi-automatic stage when iterate the panoptic model.

darkasevgen commented 6 months ago

Hello! I have a question. You wrote the following: We adopt COCO panoptic and Sama-COCO masks (high-quality instance segmentation annotations) as our starting point. But as far as I understand, in your labeling pipeline, you used in step 1 kMaX-Deeplab and DETA models to get masks and bbox proposals respectively. Hence, the COCO and Sama-COCO masks are not used, am I right?

to build the step-1 data engine, we did not use sama-coco, but used it for the step2 semi-automatic stage when iterate the panoptic model.

What do you mean? The second step in the pipeline — questionnaire

xdeng7 commented 6 months ago

Hello! I have a question. You wrote the following: We adopt COCO panoptic and Sama-COCO masks (high-quality instance segmentation annotations) as our starting point. But as far as I understand, in your labeling pipeline, you used in step 1 kMaX-Deeplab and DETA models to get masks and bbox proposals respectively. Hence, the COCO and Sama-COCO masks are not used, am I right?

to build the step-1 data engine, we did not use sama-coco, but used it for the step2 semi-automatic stage when iterate the panoptic model.

What do you mean? The second step in the pipeline — questionnaire

check 3.4 data engine, for step1, we did not use sama-coco data, all the models are trained on COCO,

darkasevgen commented 6 months ago

Hello! I have a question. You wrote the following: We adopt COCO panoptic and Sama-COCO masks (high-quality instance segmentation annotations) as our starting point. But as far as I understand, in your labeling pipeline, you used in step 1 kMaX-Deeplab and DETA models to get masks and bbox proposals respectively. Hence, the COCO and Sama-COCO masks are not used, am I right?

to build the step-1 data engine, we did not use sama-coco, but used it for the step2 semi-automatic stage when iterate the panoptic model.

What do you mean? The second step in the pipeline — questionnaire

check 3.4 data engine, for step1, we did not use sama-coco data, all the models are trained on COCO,

Then what does the quote "We adopt ... Sama-COCOCO [73] masks (high-quality instance segmentation annotations2 ) as our starting point." mean. ?