reimplement cityscapes #2089

xvjiarui · 2020-02-12T11:34:10Z

This PR reimplemented Cityscapes dataset support.

Add script to easily convert original cityscapes dataset into coco json format.
Evaluate results with Cityscapes official API.

ZwwWayne

In faster rcnn config, the type should change back to faster rcnn. Otherwise, it will cause error due to the lack of mask_head config.

The doc may indicate the users that it uses pre-trained models that need to be download from the remote servers thus the users can decide whether to download it before training or leave it for training. This might cause errors due to network or some other issues if the download is proceded in training.

ZwwWayne

Change the type back to FasterRCNN. Otherwise, it will cause bug due to the lack of config of mask_head.

The doc may indicate current configs use pre-trained models and need to download from the remote server first. Sometimes the download process might cause errors at the beginning of the training.

…to cityscapes

ZwwWayne

During implementation, I found that [7] yields better performance, maybe we can change it to [7] for higher performance and better practice.

Liyw979 · 2020-02-25T06:17:19Z

I don't think it a good idea to read Read original Cityscapes dataset annotation since it's very slow compared to using genertated coco format annotation(about 5s).
Fewer launch time is quiet important when implementing new model, during which developers frequently run trainning to see if all the tensor shapes all matched.

xvjiarui · 2020-02-25T07:41:52Z

I don't think it a good idea to read Read original Cityscapes dataset annotation since it's very slow compared to using genertated coco format annotation(about 5s).
Fewer launch time is quiet important when implementing new model, during which developers frequently run trainning to see if all the tensor shapes all matched.

Thanks for your valuable advice.

Actually, I just updated the comment of this PR. A user-friendly convert script is provided to generate COCO style annotations. You could check it out here. It would be merged into master recently.

Btw, if there is any issue about this PR, you are very welcomed to point it out or make a PR in the future.

Liyw979 · 2020-02-25T08:22:40Z

Nice! Looking forward to the new implementation:)

hellock · 2020-02-29T17:06:17Z

mmdet/datasets/cityscapes.py

+    """
+    The correct CLASSES order in official cityscapes should be
+    CLASSES = ('person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',
+                'bicycle')


This comment can be removed.

hellock · 2020-02-29T17:17:10Z

tools/convert_datasets/cityscapes.py

+    gt_dir = osp.join(cityscapes_path, args.gt_dir)
+
+    set_name = dict(
+        train='instance_train_gtFine.json',


We may also use instancesonly_filtered_gtFine_xxx.json

hellock · 2020-02-29T17:21:55Z

mmdet/datasets/cityscapes.py

+                 results,
+                 metric='bbox',
+                 logger=None,
+                 jsonfile_prefix=None,


Maybe renamed to outfile_prefix since the output file may be a json or txt.

hellock · 2020-02-29T17:22:27Z

mmdet/datasets/cityscapes.py

+        """
+        eval_results = dict()
+
+        metrics = metric if isinstance(metric, list) else [metric]


Copy it if it is a list, because we remove cityscapes from the list.

configs/cityscapes/README.md

hellock · 2020-02-29T17:28:28Z

tools/convert_datasets/cityscapes.py

+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Convert Cityscapes annotations to mmdetection format')


to COCO format

hellock · 2020-03-01T03:54:37Z

configs/cityscapes/README.md



 ### Faster R-CNN

 |    Backbone     |  Style  | Lr schd | Scale    | Mem (GB) | Train time (s/iter) | Inf time (fps) | box AP | Download |
 | :-------------: | :-----: | :-----: | :---:    | :------: | :-----------------: | :------------: | :----: | :------: |
-|    R-50-FPN     | pytorch |   1x    | 800-1024 | 4.9      | 0.345               | 8.8            | 36.0   | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/cityscapes/faster_rcnn_r50_fpn_1x_city_20190727-7b9c0534.pth) |
+|    R-50-FPN     | pytorch |   1x    | 800-1024 | 4.9      | 0.345               | 8.8            | 41.6   | [model](https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes_20200227-362cfbbf.pth) |


Train time and Inf time are outdated.

hellock · 2020-03-01T03:55:09Z

configs/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py

@@ -1,7 +1,7 @@
 # model settings
 model = dict(
    type='FasterRCNN',
-    pretrained='modelzoo://resnet50',
+    pretrained='torchvision://resnet50',


Since we have specified load_from, pretrained can be left None.

Liyw979 · 2020-03-03T15:51:22Z

The tools/test.py is running on the test set by default.

mmdetection/configs/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py

Lines 149 to 154 in 51df8a9

    
           test=dict( 
        
               type=dataset_type, 
        
               ann_file=data_root + 
        
               'annotations/instancesonly_filtered_gtFine_test.json', 
        
               img_prefix=data_root + 'leftImg8bit/test/', 
        
               pipeline=test_pipeline))

Should it be changed to val?

xvjiarui · 2020-03-03T16:04:09Z

The tools/test.py is running on the test set by default.

mmdetection/configs/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py

Lines 149 to 154 in 51df8a9

test=dict(

type=dataset_type,

ann_file=data_root +

'annotations/instancesonly_filtered_gtFine_test.json',

img_prefix=data_root + 'leftImg8bit/test/',

pipeline=test_pipeline))

Should it be changed to val?

We usually submit test set results of Cityscapes instance segmentation. You could change it to val set in your program if needed.

* reimplement cityscapes * fixed gt bbox mode * convert cityscapes to coco style, add cityscapes eval * add cityscapes convert script * add doc * Update INSTALL.md * Update INSTALL.md * update fater rcnn * fix cityscapes eval * support format only in cityscapes * add docs * remove redundancy * resolve eval * update cityscapes md * more doc and rename * update doc and cfg * change to test set

reimplement cityscapes

a7a33d3

hellock requested a review from ZwwWayne February 12, 2020 15:58

xvjiarui added 10 commits February 13, 2020 17:08

fixed gt bbox mode

52df1bd

convert cityscapes to coco style, add cityscapes eval

3554bc1

merge master fixed conflict

5deed7a

add cityscapes convert script

258a219

add doc

c5790e2

Update INSTALL.md

c941147

Update INSTALL.md

d13a652

Merge branch 'master' into cityscapes

25589f3

update fater rcnn

c49c120

fix cityscapes eval

caefc42

xvjiarui force-pushed the cityscapes branch from 319216f to caefc42 Compare February 17, 2020 17:45

support format only in cityscapes

4ddce4e

ZwwWayne reviewed Feb 21, 2020

View reviewed changes

xvjiarui added 4 commits February 21, 2020 17:04

add docs

d67d7d5

Merge branch 'master' into cityscapes

446aae1

Merge branch 'master' of https://github.com/open-mmlab/mmdetection in…

74e1419

…to cityscapes

remove redundancy

4220316

ZwwWayne reviewed Feb 24, 2020

View reviewed changes

xvjiarui added 2 commits February 25, 2020 16:54

resolve eval

72cbcf2

update cityscapes md

a25a692

hellock mentioned this pull request Feb 28, 2020

Fix cityscapes class order #1758

Closed

hellock reviewed Feb 29, 2020

View reviewed changes

more doc and rename

eee60b3

hellock reviewed Mar 1, 2020

View reviewed changes

xvjiarui added 4 commits March 1, 2020 13:58

update doc and cfg

7184a45

Merge branch 'master' into cityscapes

0b11bcc

change to test set

7eb5600

resolve conflict

2eed951

hellock approved these changes Mar 1, 2020

View reviewed changes

hellock merged commit e27046d into open-mmlab:master Mar 1, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

reimplement cityscapes #2089

reimplement cityscapes #2089

xvjiarui commented Feb 12, 2020 •

edited

Loading

ZwwWayne left a comment

ZwwWayne left a comment

ZwwWayne left a comment

Liyw979 commented Feb 25, 2020 •

edited

Loading

xvjiarui commented Feb 25, 2020

Liyw979 commented Feb 25, 2020

hellock Feb 29, 2020

hellock Feb 29, 2020

hellock Feb 29, 2020

hellock Feb 29, 2020

hellock Feb 29, 2020

hellock Mar 1, 2020

hellock Mar 1, 2020

Liyw979 commented Mar 3, 2020

xvjiarui commented Mar 3, 2020

reimplement cityscapes #2089

reimplement cityscapes #2089

Conversation

xvjiarui commented Feb 12, 2020 • edited Loading

ZwwWayne left a comment

Choose a reason for hiding this comment

ZwwWayne left a comment

Choose a reason for hiding this comment

ZwwWayne left a comment

Choose a reason for hiding this comment

Liyw979 commented Feb 25, 2020 • edited Loading

xvjiarui commented Feb 25, 2020

Liyw979 commented Feb 25, 2020

hellock Feb 29, 2020

Choose a reason for hiding this comment

hellock Feb 29, 2020

Choose a reason for hiding this comment

hellock Feb 29, 2020

Choose a reason for hiding this comment

hellock Feb 29, 2020

Choose a reason for hiding this comment

hellock Feb 29, 2020

Choose a reason for hiding this comment

hellock Mar 1, 2020

Choose a reason for hiding this comment

hellock Mar 1, 2020

Choose a reason for hiding this comment

Liyw979 commented Mar 3, 2020

xvjiarui commented Mar 3, 2020

xvjiarui commented Feb 12, 2020 •

edited

Loading

Liyw979 commented Feb 25, 2020 •

edited

Loading