How to improve FPS for OBB detection on customized dataset ?

[starlord93]

Hi Wang, I tried three different OBB models (gliding_vertex_r50_fpn_1x, faster_rcnn_orpn_r50_fpn_1x and retinanet_obb_r50_fpn_1x) on customized dataset adapted from PCB Dataset. In average, the mAP scores are ~90% for all 3 models.However, the FPS or tasks/seconds is only ~3. The test images are resized to 1024*1024 and the FPS data were obtained from tools/test.py (I believe is only single_gpu_testing).

One of the result obtained on test images: [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 75/75, 2.7 task/s, elapsed: 28s, ETA: 0s +-------+-----+------+--------+--------+ | class | gts | dets | recall | ap | +-------+-----+------+--------+--------+ | ic | 729 | 946 | 0.9767 | 0.8992 | +-------+-----+------+--------+--------+ | mAP | | | | 0.8992 | +-------+-----+------+--------+--------+

I did not split the images during train & test. Can splitting improve the FPS ? Or multi GPU can improve the FPS ?

I attached some files for more details of my pipelines and models. Looking forward to your reply. Thanks pcb.txt gliding_vertex_r50_fpn_1x_pcb.txt

[jbwang1997]

The multi GPU testing can improve the FPS. Also, you can also resize images to a smaller size like 600 600. The 1024 1024 image may be a little big.

[wkmwkm93]

I see. Is there any guide how to edit the tools/dist_test.sh file ? Does it mean that training images also need to be the same size as testing image sizes ?

[jbwang1997]

dist_test.sh also runs the test.py. So you can edit the test.py for your task.

Yes, the training images and testing images should be the same size.

[wkmwkm93]

I got this error when I executed: %run tools/test.py configs/obb/gliding_vertex/gliding_vertex_r50_fpn_1x_pcb.py gv_eval_output/epoch_11.pth --eval mAP --launcher pytorch

/usr/local/lib/python3.7/dist-packages/mmcv/runner/dist_utils.py in init_dist(launcher, backend, kwargs) 15 mp.set_start_method('spawn') 16 if launcher == 'pytorch': ---> 17 _init_dist_pytorch(backend, kwargs) 18 elif launcher == 'mpi': 19 _init_dist_mpi(backend, **kwargs)

/usr/local/lib/python3.7/dist-packages/mmcv/runner/dist_utils.py in _init_dist_pytorch(backend, kwargs) 26 def _init_dist_pytorch(backend, kwargs): 27 # TODO: use local_rank instead of rank % num_gpus ---> 28 rank = int(os.environ['RANK']) 29 num_gpus = torch.cuda.device_count() 30 torch.cuda.set_device(rank % num_gpus)

/usr/lib/python3.7/os.py in getitem(self, key) 679 except KeyError: 680 # raise KeyError with the original key value --> 681 raise KeyError(key) from None 682 return self.decodevalue(value) 683

KeyError: 'RANK'