This reverts commit 47fc188b.

* allow forwarding for multi scale inputs

* fixed PadBatch not working with multi scale input

* add faster_rcnn_r34_fpn_multiscaletest_1x_coco.yml

* fixed multi scale input checking for batch transforms (PadBatch)

* less scales and faster inference for multi scale test configuration file

* fixed retrieving im_id from multi-scale inputs

* add multi scale prediction boxes merging

* allow multi scale testing for tools/infer.py (actually trainer.predict method)

* add MultiscaleTestResize setting for TestReader

* use BBoxPostProcess.nms parameter for multi scale prediction NMS of CascadeRCNN

* add documentation for multi scale testing

* raise exception if the network architecture is not supported in multi scale test mode

* move documentation file

* add test case for multi scale test

* load pretrained weights from URL

* add image files for testing. use relative path for input images

* update documentation

* use isinstance for sequence type checking. add empty checking for samples sequence

* fix operators typo

* fix kitti metric deploy

* modify VOCDataSet and default value of allow_empty

* revert default value of allow_empty (#4150)
Co-authored-by: Nwangguanzhong <jerrywgz@126.com>

* fix operators typo

* fix centernet_head bias init

* fix operators typo

* fix mot typo, test=document_fix

* fix operators typo

* add BasicBlock dcn

* add frame_rate for mot infer video

* add doc of frame_rate for mot video infer, test=document_fix

* [dev] rbox update2 (#3828)

* set lr for 4 card as default, and update

* fix error

* add kitti metric

* clean kitti metric code

* fix kitticars doc

* fix feature model cfgs

* add kitti metric

* clean code of s2anet

* fix for paddle==2.1.0

* remove debug info

* update comment

* mot video decode to images

* add infer_dir for mot

* fix export and unite pose infer

* remove task in mot source

* set image_dir

* update doc