Segmentation training issues: StopIteration

[yikakabu]

HI! Thank you for the great job. I created a json file of my own annotated data by following the coco mask.py. The part of the json file as below:

{..., "image": [{"license": 0, "file_name": "0000007149_rgb.png", "coco_url": "", "height": 480, "width": 640, "date_captured": 1544011401.0, "camera_id": 0, "flickr_url": "", "id": 0},...], "annotations": [{"segmentation": {"size": [480, 640], "counts": "QgY37h>f0ZOe0\Od0\Od0\Od0[Of0[Oc0]O`0@3M2N2N3M2N2N2N3L3N2N2N3M2N2N2N2N1O001O0O2O0000001O000000O1000O10000000O10000O100000O0100000000O100000000O1000O10O1000000O100000000O0100000O100000000O1000O10O1000000O1000000O1000O1000O1000000O1000000O10O100000O1000000O10000000O0100000000O1000000O1000O10O100000000O100000O0100000000O10000O100000001N1000001N10001O0O101O000O2O0000001N101hMfFmLn0a0m;]OUDa0U=01N2O1O1O1N101O1O1N2O1O1N2N2MQeW3"}, "area": 34310, "pose": {"position": {"x": -0.0007623627074502259, "y": 0.34552469760243687, "z": 1.0687215939143497}, "orientation": {"x": 0.7472581634432386, "y": -0.14419430357527752, "z": 0.10985900520900949, "w": 0.6393310871202543}}, "iscrowd": 0, "image_id": 0, "bbox": [225.0, 242.0, 193.0, 219.0], "category_id": 1, "id": 0},...], "categories": [{"supercategory": "objects", "mesh": "charge_pile.stl", "id": 1, "name": "charge_pile"}]}

Then I modified the config file(drn_d_22_ChargePile.yml) and utils.py related to the object classes in "segmentation" folder. But when I run

python train.py config/drn_d_22_ChargePile.yml

There is a error "StopIteration" came out at: first_input_batch, first_target_batch = iter(val_loader).next() in the train.py:217 and RuntimeWarning: invalid value encountered in true_divide at: return np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist)) in the utils.py:25 It seems my dataset wasn't loaded correctly. Can you give me any suggestions about how to solve these problem？Thanks.

[jimmyyhwu]

Maybe you could use matplotlib to visualize the images that are being loaded by val_loader?

The StopIteration error probably means val_loader isn't loading any images. I would check that self.img_ids in datasets.py is correct.

[yikakabu]

@jimmyyhwu Thanks for your reply. You are right. That is because my annotations in the json file had wrong format. I have already solved the issue and can start training now.

[yikakabu]

@jimmyyhwu Could you please explain the meaning of the parameter "object_scales" in generate_pose_lists.py. And how can I get the parameter from my own CAD model? Thanks very much.

object_scales = {
        1: 0.3,
        2: 0.3,
        3: 0.2,
        4: 0.2,
        5: 0.7,
        6: 0.2,
        7: 0.2,
        8: 0.3,
        9: 0.3,
        10: 0.2,
    }

[jimmyyhwu]

The object_scale parameters are used to determine the position mean/cov when generating random positions here. Larger objects are usually further away from the camera. For example, the engine (scale=0.7) is much larger than the oil filter (scale=0.2). See object IDs here.

You will probably need to visually verify your rendered images and make sure the objects aren't too close or too far away.

[yikakabu]

@jimmyyhwu Thanks for your reply. Dose that means a appropriate distance from the object to camera? So, I should determine the scale of my model according to the size of the object. Do I understand that correctly?

[jimmyyhwu]

Yes, it should roughly be proportional to the size of object. You can visually verify that the objects have appropriate sizes in the rendered images.

By the way, for further follow-up on this, it would be good to open a new, separate issue, since the object_scale is not relevant to the "StopIteration" error originally referenced in this issue.

[yikakabu]

@jimmyyhwu I have understanded that. Thank you.