diff --git a/README.md b/README.md index 5b51ff3..088cef5 100644 --- a/README.md +++ b/README.md @@ -5,6 +5,22 @@ This repo uses Centernet and Conditional Convolutions for Instance Segmentation > [**Objects as Points**](http://arxiv.org/abs/1904.07850), > [**CondInst: Conditional Convolutions for Instance Segmentation**](https://arxiv.org/abs/2003.05664) +## Result + +These results are taken for CenterSeg model trained for 101 epochs + +| type | AP | AP50 | AP75 | APs | APm | APl | +| ---- | ----- | --------------- | --------------- | -------------- | -------------- | -------------- | +| box | 0.278 | 0.430 | 0.297 | 0.129 | 0.305 | 0.382 | +| mask | 0.226 | 0.387 | 0.227 | 0.078 | 0.253 | 0.340 | + +| type | AR | AR50 | AR75 | ARs | ARm | ARl | +| ---- | ----- | --------------- | --------------- | -------------- | -------------- | -------------- | +| box | 0.275 | 0.455 | 0.480 | 0.265 | 0.510 | 0.674 | +| mask | 0.235 | 0.369 | 0.385 | 0.170 | 0.418 | 0.585 | + +CenterPoseSeg model not trained yet + ## Installation This repo supports both CPU and GPU Training and Inference. @@ -24,6 +40,7 @@ python3 setup.py build develop ``` Compile NMS + ``` cd src/lib/external @@ -47,16 +64,19 @@ Note: Model is not completely trained (101 Epochs only). Will update later. #### Training ###### For GPU + ``` python3 main.py ctseg --exp_id coco_dla_1x --batch_size 10 --master_batch 5 --lr 1.25e-4 --gpus 0 --num_workers 4 ``` ###### FOR CPU + ``` python3 main.py ctseg --exp_id coco_dla_1x --batch_size 2 --master_batch -1 --lr 1.25e-4 --gpus -1 --num_workers 4 ``` #### Testing + ``` python3 test.py ctseg --exp_id coco_dla_1x --keep_res --resume ``` @@ -73,4 +93,3 @@ https://github.com/xingyizhou https://github.com/Epiphqny https://github.com/CaoWGG - diff --git a/src/lib/trains/ctseg.py b/src/lib/trains/ctseg.py index d5721fc..85f1dda 100644 --- a/src/lib/trains/ctseg.py +++ b/src/lib/trains/ctseg.py @@ -68,7 +68,6 @@ def forward(self, outputs, batch): if opt.reg_offset and opt.off_weight > 0: off_loss += self.crit_reg(output['reg'], batch['reg_mask'], batch['ind'], batch['reg']) / opt.num_stacks - mask_loss += self.crit_mask(output['seg_feat'], output['conv_weight'], batch['reg_mask'], batch['ind'], batch['instance_mask'])