Bin Yan

@iminfine Hi, we train STARK with 8x 16GB Tesla V100 GPUs and the training does only takes about 2 days. This has also been verified by other teams. The extremely...

Hi, after converting the original PyTorch model to the ONNX model, you don't need to further convert it to TensorRT model. Actually, onnxruntime has tensorrt execution provider [(ONNXRUNTIME/execution-providers/TensorRT).](https://onnxruntime.ai/docs/reference/execution-providers/TensorRT-ExecutionProvider.html#python) But to...

@trathpai Hi, thanks for the appreciation of our work. ONNXRuntime has a TensorRT execution provider. This link (https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html) may give you some help. But from my experience, TensorRT execution provider...

@orilifs Hi, thanks for your appreciation of our work. Answer to Q1.a: These two models are part of our improved version of STARK (we call it STARK-Lightning). The first one...

@elvindp Hi, the mask is used to deal with padding. We don't want the network (Transformer) to learn similarity with padding regions because these regions have no valuable information. The...

@JuiceLemonLemon Hi, we found that the default tolerance (relative 1e-3, absolute 1e-5) may not be suitable for our network outputs. Specifically, the testing results (whether it can pass the testing)...

@ANdong-star Hi, this process is quite similar to that in the DETR decoder. In DETR, 100 object queries interact with the image features output by the encoder. In STARK, one...

Hi, thanks for your appreciation in our work. I think that replacing the original transformer with Swin-T may reduce some computational complexity but will not bring performance gain, because the...

@nubsym Hi, thanks for your appreciation of our work. You can refer to for the detailed training log of STARK.

Hi, of course. Could you tell me the detailed config name (baseline / baseline_got10k_only / baseline_R101 / baseline_R101_got10k_only) ? Or I can upload them together to the model zoo. Thanks.