对象检测tensorflow api中的ssd mobilenet v1可以尝试使用不同于默认值的调整大小形状吗？

Question

In ssd_mobilenet_v1_coco.config the image_resizer default size is 300x300, or 512x512. State of art results are available for the options only.

But resizing to smaller sizes leads to information loss, can ssd mobilenet be tried with say size 720x720?

Config file: https://github.com/tensorflow/models/blob/master/research/object_detection/samples/configs/ssd_mobilenet_v1_coco.config

Answer 1

It depends whether you're asking about training or inference.

If your goal is to detect objects using a pre-trained model, then it is not recommended to change the resizing parameters, as the model is tuned to work best of these.

However, if you wish to train the model, then yes, you can modify them. However, be aware that changing these values non-marginally would also require you to change the architecture and/or anchor configuration a bit, depending on the objects' sizes you wish to detect. For example, if you're using larger input resolution, than I would recommend adding SSD layers ( this is the original configuration, with 6 feature maps with stride of 8, 16, 32, 64, 128 and 256) and changing anchor scales ( this is the original, with 6 layers and linear scales in the range of 0.2-0.95 of the image input size).