about SOTR-RT-736

[Anglechina]

hi, i can't find the model about SOTR-RT-736 to test pic at high FPS, when i am reshowing your work, can you give me some help? 1) can you provide the model about SOTR-RT-736? 2) can you tell me some details about test the model ? thank you very much!

[easton-cau]

SOTR mainly focus on boosting the accuracy of instance segmentation. When it comes to real-time models, SOTR-RT has lower accuracy and speed compared with other state-of-the-art instance segmentation methods. So we recommend you choose other CNN methods if your task is a real-time task. If you want to test the SOTR-RT model, you can reduce the number of transformer layers to two and the input shorter side to 736, and then train the model.

[Anglechina]

SOTR mainly focus on boosting the accuracy of instance segmentation. When it comes to real-time models, SOTR-RT has lower accuracy and speed compared with other state-of-the-art instance segmentation methods. So we recommend you choose other CNN methods if your task is a real-time task. If you want to test the SOTR-RT model, you can reduce the number of transformer layers to two and the input shorter side to 736, and then train the model.

when i test the SOTR_R101 or SOTR_R101DCN model, the cost time is about 0.2~0.3s using GPU3080, is that correct comparing with your work?

[easton-cau]

We test the SOTR_R101 model on COCO test-dev, and the cost time is about 0.123 s/per using a single Tesla V100 GPU.

Please check the following reasons: 1) Image resolution. 2) Function. demo.py: demo.run_on_image(img) predictor.py: visualizer.draw_instance_predictions(predictions=instances) # need lots of time when drawing predicted pictures. We chose to directly generate json file instead of pictures. 3) GPU. Tesla V100 vs. GTX 3080