Do you report MSCOCO performance? #1

gaopeng-eugene · 2017-11-01T23:55:05Z

No description provided.

waleedka · 2017-11-02T01:00:15Z

Evaluation code against MS COCO is included in the repository, both for bounding boxes and segmentation masks so it should be easy to run (but takes a long time).

We should publish more details, though. Thanks for bringing it up. Our implementation deviates a bit from the paper (as mentioned in the documentation), and optimizing for COCO was a 'nice to have' rather than being the main objective. We got pretty close to the reported numbers (within 3 to 4 percentage points) but that was with half the training steps compared to the paper. We'll try to add more details over the next few days.

larsoncs · 2017-11-08T07:17:18Z

I used the pre-trained COCO weights test，After testing segmentation masks, But i found the result is not good ,could you tell me the reason?

yanxp · 2017-11-09T07:47:32Z

I trained the coco dataset and found the result was not good . How to fine tune the work or set the parameters,why is that?

ppwwyyxx · 2017-11-09T18:31:32Z

Out of curiosity, what's the score people have seen on COCO?

vijaykbg · 2017-11-10T05:10:41Z

waleedka · 2017-11-10T06:20:26Z

@vijaykbg That's a lot lower than what I got. Is this for bounding boxes or segmentation masks? I'll try to generate and share our results next week.

larsoncs · 2017-11-10T06:25:24Z

@waleedka it's great work，looking forward to your next week results and model weights file !!!!

vijaykbg · 2017-11-10T08:15:36Z

@waleedka Thanks for the info and the code. I evaluated segmentation masks ("segm") using the evaluation code coco.py.
It would be really helpful if you could share the updated models and your results soon.
Thanks in advance!

csnemes2 · 2017-11-13T13:35:26Z

I have installed this right now, and without any tweak I run

python3 coco.py evaluate --dataset=/home/csn/COCO/ --model=./mask_rcnn_coco.h5

Which probably translates to

   # TODO: evaluating on 500 images. Set to 0 to evaluate on all images.
   evaluate_coco(model, dataset_val, coco, "bbox", limit=500)

Which finally gives me the following precisions:

Evaluate annotation type bbox
DONE (t=1.09s).
Accumulating evaluation results...
DONE (t=0.34s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.298
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.466
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.332
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.133
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.350
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.435
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.248
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.344
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.351
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.146
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.407
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.508

csnemes2 · 2017-11-13T14:04:21Z

And for "segm":

Evaluate annotation type *segm*
DONE (t=1.20s).
Accumulating evaluation results...
DONE (t=0.34s).
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.267
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.453
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.283
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.106
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.313
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.410
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.225
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.309
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.316
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.119
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.365
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.466

Walid-Ahmed · 2017-11-13T18:58:13Z

@csnemes2
How can I actually run "evaluate" for segm ?
Thanks

ypflll · 2017-11-17T02:43:30Z

@Walid-Ahmed
Do u mean this:
evaluate_coco(model, dataset_val, coco, "segm", limit=500)?

ypflll · 2017-11-17T02:44:00Z

Evaluation is really time-consuming and seems that my gpu is not used, when configurationis like this:
Configurations:
BACKBONE_SHAPES [[256 256]
[128 128]
[ 64 64]
[ 32 32]
[ 16 16]]
BACKBONE_STRIDES [4, 8, 16, 32, 64]
BATCH_SIZE 1
BBOX_STD_DEV [ 0.1 0.1 0.2 0.2]
DETECTION_MAX_INSTANCES 100
DETECTION_MIN_CONFIDENCE 0.7
DETECTION_NMS_THRESHOLD 0.3
GPU_COUNT 1
IMAGES_PER_GPU 1
......

@waleedka How can I change it to evalute on gpu, not cpu?

ypflll · 2017-11-17T06:35:59Z

Just falsely installed the cpu version of tensorflow.
With gpu version, 10 tmies less time cost^_^.

jiang1st · 2017-11-20T09:06:05Z

Has anyone reproduced the result published on the original paper? I tried the code, but the model is converging slowly. The mAP is around 29%.
Update: I evaluated the model on minival5k.

ppwwyyxx · 2017-11-20T12:24:13Z

The pretrained model seems to base on ResNet101-FPN backbone if I'm not mistaken. The reference number in the paper for this architecture is segm-~~32.4~~35.4 (on minival), bbox-38.2 (on test-dev). According to @csnemes2 's evaluation the number here is segm-26.7, bbox-29.8. This is a large gap IMHO, not 3 to 4 percentage points.

jiang1st · 2017-11-21T02:04:54Z

@ppwwyyxx It seems @csnemes2 only evaluated the top500 images.

ypflll · 2017-11-21T03:59:11Z

@ppwwyyxx I see in Kaiming He's paper, for ResNet-101-FPN backbone, AP is 35.4 on 5k minival and 35.7 on COCO test-dev.
@jiang1st I just tested on minival, AP is 0.261 for segm and 02.98 for bbox.

@waleedka Hi, since we can't reproduce the results, can you post your results on coco? Thanks.

csnemes2 · 2017-11-21T08:20:58Z

@ppwwyyxx True

Usually there is no big difference, hence, the success/popularity of minival

Anyway I started a new evaluation on the whole dataset, but it will take a while

waleedka · 2017-11-26T04:49:42Z

This took a bit longer than I hoped, but I just uploaded a new trained model.

https://github.com/matterport/Mask_RCNN/releases/tag/v2.0

These are the results on the COCO 5K minival. Bounding box AP is 34.7 (the paper reports 38.2 on test-dev but they don't report minival, unless I missed it). For segmentation, the AP is 29.6 (the paper reports 35.4 on minival). Playing with hyper parameters might allow you to get higher accuracy, but I didn't have enough time to try.

Evaluate annotation type *bbox*
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.347
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.544
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.377
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.163
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.390
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.486
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.295
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.424
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.433
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.214
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.481
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.601

Evaluate annotation type *segm*
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.296
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.510
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.306
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.128
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.330
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.430
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.258
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.369
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.376
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.173
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.417
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.538

ppwwyyxx · 2017-11-26T06:58:21Z

FYI I pushed my own implementation. On minival it gets 40.1 on bbox and 34.4 on segm, with only a ResNet101 backbone (no FPN). This is slightly better than the paper's number.

jiang1st · 2017-11-26T10:48:41Z

@waleedka Thank you for your efforts. Another thing I want to mention is, why are you loading parameters from a resnet50-based pretrained model (see model.get_imagenet_weights()) when training for a resnet101-based network?
One more question, when you set learning rate to 0.001, what is the actual batch size you are using? 16 (two image by 8 GPU) ? Thanks.

waleedka · 2017-11-27T00:10:53Z

@ppwwyyxx That's awesome. Do you think your custom implementation of crop_and_resize is the main factor in your performance? I've seen other implementations that use the standard crop_and_resize and still get very good accuracy. So I'm curious how big a difference that contributed to your results.

waleedka · 2017-11-27T00:35:26Z

@jiang1st Keras provides a resnet50 trained weights, but not resnet101. I could find resnet101 weights from another source and do a quick mapping of weight names and load them. I just didn't get around to doing so. The residual connections in resnet make this less of an issue, other than longer fine tuning time, and for most people starting from the COCO weights is a faster path. Still, I do want to add 101 weights when I get a chance.

Regarding the learning rate, I'm using 0.001 with a batch size of 16 (2 images per GPU on 8 GPUs). Typically you'd want to use a higher LR with bigger batches, but my observation in a few tests lead me to believe that the gradients from different GPUs are being summed (rather than averaged). This is a guess as I haven't verified it by tracing the internal implementation code, but if it's true then it means that you could use the same LR with regardless of the batch size.

ppwwyyxx · 2017-11-27T03:08:15Z

@waleedka It improves a little bit (within 1 mAP). If your model is not initialized with a pretrained resnet101, it is probably the main reason.

waleedka · 2017-11-27T07:42:20Z

@ppwwyyxx That makes sense, thanks! I'll train on resnet101 weights and see how much improvement that gives me. And if I get the chance, I want to try your implementation of ROI Align and see if that gives it additional boost as well.

jmtatsch · 2017-11-27T13:44:05Z

@waleedka there is converted resnet101 weights for keras here https://gist.github.com/flyyufelix/65018873f8cb2bbe95f429c474aa1294
they load just fine when you set use_bias=False in conv1, conv_block & identity block. Apparently thats ok because the bias should be in the bn layers anyway KaimingHe/deep-residual-networks#10

jmtatsch · 2017-12-10T11:52:42Z

@waleedka I trained with your latest training schedule on a 4 gpu system but get far worse results than your v2 weights. How exactly did you train your v2 weights: starting from imagenet? Same as your published training schedule? Is the difference between a 4 and a 8 gpu system and the resulting larger batch sizes really that big?

jiang1st · 2017-12-11T11:56:18Z

@jmtatsch I tried with 8 gpus with the default configuration, and didn't reproduce the reported result (far below 34.7). Not sure where could be the problem.

John1231983 · 2018-02-02T18:54:31Z

As the @ppwwyyxx said, the problem may come from pretrained model. So why not write resnet in model.py by using tf-slim and then you can use tf-slim official pretrained model. Is it possible?

John1231983 · 2018-02-02T19:43:18Z

What does it mean of full pretrained model in compare with resnet pretrain?

John1231983 · 2018-02-02T21:21:27Z

So my question is that can we rewrite resnet part using tf-slim in model.py to use offical pretrained model in tf-slim team

John1231983 · 2018-02-03T05:33:39Z

I see what you said. So, we first convert pretrained resnet 101 in tf-slim to keras and we can train with full pretrained model.

jmtatsch · 2018-02-05T10:43:03Z

@XupingZHENG With the 101 layer weights and waleeds training schedule i got very mediocre results. Pretty much doubled the training up and its still running.

endluo · 2018-07-18T08:27:20Z

i train the coco2017, the loss is 0.6 , but i get the result is zeros.I don't know why...
this is my command:CUDA_VISIBLE_DEVICES=1 python3 coco.py evaluate --dataset=/home/swland/coco/ --model=/home/swland/coco/mask_rcnn_coco_last.h5 --year=2017

Configurations:
BACKBONE resnet101
BACKBONE_STRIDES [4, 8, 16, 32, 64]
BATCH_SIZE 1
BBOX_STD_DEV [0.1 0.1 0.2 0.2]
COMPUTE_BACKBONE_SHAPE None
DETECTION_MAX_INSTANCES 100
DETECTION_MIN_CONFIDENCE 0
DETECTION_NMS_THRESHOLD 0.3
FPN_CLASSIF_FC_LAYERS_SIZE 1024
GPU_COUNT 1
GRADIENT_CLIP_NORM 5.0
IMAGES_PER_GPU 1
IMAGE_MAX_DIM 512
IMAGE_META_SIZE 18
IMAGE_MIN_DIM 512
IMAGE_MIN_SCALE 1
IMAGE_RESIZE_MODE square
IMAGE_SHAPE [512 512 3]
LEARNING_MOMENTUM 0.9
LEARNING_RATE 0.001
LOSS_WEIGHTS {'rpn_bbox_loss': 1.0, 'mrcnn_class_loss': 1.0, 'mrcnn_mask_loss': 1.0, 'rpn_class_loss': 1.0, 'mrcnn_bbox_loss': 1.0}
MASK_POOL_SIZE 14
MASK_SHAPE [28, 28]
MAX_GT_INSTANCES 100
MEAN_PIXEL [123.7 116.8 103.9]
MINI_MASK_SHAPE (56, 56)
NAME coco
NUM_CLASSES 6
POOL_SIZE 7
POST_NMS_ROIS_INFERENCE 1000
POST_NMS_ROIS_TRAINING 2000
ROI_POSITIVE_RATIO 0.33
RPN_ANCHOR_RATIOS [0.5, 1, 2]
RPN_ANCHOR_SCALES (32, 64, 128, 256, 512)
RPN_ANCHOR_STRIDE 1
RPN_BBOX_STD_DEV [0.1 0.1 0.2 0.2]
RPN_NMS_THRESHOLD 0.7
RPN_TRAIN_ANCHORS_PER_IMAGE 256
STEPS_PER_EPOCH 1000
TOP_DOWN_PYRAMID_SIZE 256
TRAIN_BN False
TRAIN_ROIS_PER_IMAGE 200
USE_MINI_MASK True
USE_RPN_ROIS True
VALIDATION_STEPS 50
WEIGHT_DECAY 0.0001

Running per image evaluation...
Evaluate annotation type bbox
DONE (t=6.00s).
Accumulating evaluation results...
DONE (t=1.08s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
Prediction time: 315.4280889034271. Average 0.6308561778068542/image
Total time: 428.98461985588074

michaelisc · 2018-08-06T11:25:05Z

@waleedka Could you post the hyperparameters you used to get the reported performance somewhere? The results when I train the model look way worse and with a model as complex as Mask R-CNN doing a hyperparameter search takes a lot of time. So it would be great to know, which configuration you used for the provided checkpoint and the posted results.

s-bayer · 2018-08-08T10:13:59Z

Any updates on this? I'm trying to use this repository as baseline for my thesis, which is impossible if results are so far off from the original paper.
I'm also currently running experiments and will report if I find a way to make it work, but as said above: experimenting on this is quite expensive and takes a lot of time.
@waleedka, I would really appreciate it, if you could post your exact training code, so that I can use it :)

s-bayer · 2018-08-13T19:49:52Z

I adjusted the training schedule to something more reasonable and am slowly getting somewhere. Currently bbox-AP is at 31.4 and segm-AP at 29.1 with a Resnet-101 on minival2014.

Still not great, but way better than what I got with the training schedule posted with this repository.
My training is still running and I should have final results in 1-2 days and will keep you updated.

I will of course post my training schedule eventually, but would like to wait until my thesis is done (October 1). If someone needs it earlier, just message me.

michaelisc · 2018-08-13T21:08:04Z

@syrix I would actually be very interested. I've also updated my schedule and get better results now but they are still quite a bit from the official checkpoint (however using a ResNet50).

s-bayer · 2018-08-14T19:23:44Z

Training is done now. Final results are 32.6 for bbox and 30.1 for segm on 8-GPUs with a Resnet101 initialized with Resnet50 weights. There are also checkpoints with slightly better numbers (33.0 and 30.4), but I will just use the last one.

Not sure where the higher values from @waleedka come from, but I'm fairly sure it's not just changes to the learning schedule, but changes to the remaining code, e.g. not freezing batchnorm.

I use a base learning rate of 0.005 (which is fairly aggressive) and a batch size of 2 for an effective batch size of 16.
My schedule is as follows (epochs are 1000 steps each and you need to add all epochs to get the total number of epochs):

    trainer = Trainer(model, dataset_train, dataset_val, config, augmentation)

    trainer.train('Training Stage 1: only network heads',
                  num_epochs=1, layers='heads')

    trainer.train('Training Stage 2: warmup for network except resnet-50',
                  num_epochs=1, layers='4+', learning_rate_factor=0.01)

    trainer.train('Training Stage 3: train network except resnet-50',
                  num_epochs=5, layers='4+')

    trainer.train('Training Stage 4: warmup for everything',
                  num_epochs=1, layers='all', learning_rate_factor=0.01)

    trainer.train('Training Stage 5: train everything',
                  num_epochs=100, layers='all')

    trainer.train('Training Stage 6: fine-tune / 10',
                  num_epochs=20, layers='all', learning_rate_factor=0.1)

    trainer.train('Training Stage 7: fine-tune / 100',
                  num_epochs=10, layers='all', learning_rate_factor=0.01)

The final fine-tuning / 100 barely changes results, so 1-2 epochs of that is probably enough.
You could also remove around 20 or maybe even 30 epochs of stage 5, if you need to save on training time.

The last three stages are intentionally fairly similar to the s1x-schedule of Detectron, which is from the authors of the paper, but my learning rate is different and I don't use linear warmup.

Here is the trainer-class in case you need it:

class Trainer:
    def __init__(self, model, dataset_train, dataset_val, config, augmentation):
        self.model = model
        self.dataset_train = dataset_train
        self.dataset_val = dataset_val
        self.config = config
        self.augmentation = augmentation
        self.total_epochs = 0

    def train(self, message, num_epochs, layers, learning_rate_factor=1.0):
        self.total_epochs += num_epochs
        print(message + f' for {num_epochs} mini-epochs.')
        self.model.train(self.dataset_train, self.dataset_val,
                         learning_rate=self.config.LEARNING_RATE * learning_rate_factor,
                         epochs=self.total_epochs,
                         layers=layers,
                         augmentation=self.augmentation)

And here are loss curves.

Train-loss:

Val-loss:

You can clearly see that the first learning rate decrease is essential, while the second one is not necessary.

stark-xf · 2018-10-17T07:44:49Z

I train the coco2017, use coco.py, but get the result mAP=0.203. It is far worse result, could you tell me the reason?

Command: train
Model: /home/lixiaofeng/project/version5/Mask_RCNN-master/logs/coco20181012T1737/mask_rcnn_coco_0020.h5
Dataset: /home/lixiaofeng/dataset/coco
Year: 2017
Logs: /home/lixiaofeng/project/version5/Mask_RCNN-master/logs
Auto Download: False

Configurations:
BACKBONE resnet101
BACKBONE_STRIDES [4, 8, 16, 32, 64]
BATCH_SIZE 4
BBOX_STD_DEV [0.1 0.1 0.2 0.2]
COMPUTE_BACKBONE_SHAPE None
DETECTION_MAX_INSTANCES 100
DETECTION_MIN_CONFIDENCE 0.7
DETECTION_NMS_THRESHOLD 0.3
FPN_CLASSIF_FC_LAYERS_SIZE 1024
GPU_COUNT 4
GRADIENT_CLIP_NORM 5.0
IMAGES_PER_GPU 1
IMAGE_MAX_DIM 1024
IMAGE_META_SIZE 93
IMAGE_MIN_DIM 800
IMAGE_MIN_SCALE 0
IMAGE_RESIZE_MODE square
IMAGE_SHAPE [1024 1024 3]
LEARNING_MOMENTUM 0.9
LEARNING_RATE 0.001
LOSS_WEIGHTS {'rpn_class_loss': 1.0, 'rpn_bbox_loss': 1.0, 'mrcnn_class_loss': 1.0, 'mrcnn_bbox_loss': 1.0, 'mrcnn_mask_loss': 1.0}
MASK_POOL_SIZE 14
MASK_SHAPE [28, 28]
MAX_GT_INSTANCES 100
MEAN_PIXEL [123.7 116.8 103.9]
MINI_MASK_SHAPE (56, 56)
NAME coco
NUM_CLASSES 81
POOL_SIZE 7
POST_NMS_ROIS_INFERENCE 1000
POST_NMS_ROIS_TRAINING 2000
ROI_POSITIVE_RATIO 0.33
RPN_ANCHOR_RATIOS [0.5, 1, 2]
RPN_ANCHOR_SCALES (32, 64, 128, 256, 512)
RPN_ANCHOR_STRIDE 1
RPN_BBOX_STD_DEV [0.1 0.1 0.2 0.2]
RPN_NMS_THRESHOLD 0.7
RPN_TRAIN_ANCHORS_PER_IMAGE 256
STEPS_PER_EPOCH 1000
TOP_DOWN_PYRAMID_SIZE 256
TRAIN_BN False
TRAIN_ROIS_PER_IMAGE 200
USE_MINI_MASK True
USE_RPN_ROIS True
VALIDATION_STEPS 50
WEIGHT_DECAY 0.0001

SG-97 · 2020-09-07T14:51:30Z

I'm running the evaluation but it seems to be stuck at "index created!" then, nothing. I've waited for 12+, evaluating on just 100 images. Yet the results aren't ready ? Is this an error of some sort or should just preserver.

Bugfix for MaskRCNN creating empty log_dir that breaks find_last()

Tensorflow2.0

ashwin-999 · 2021-06-22T22:08:59Z

This took a bit longer than I hoped, but I just uploaded a new trained model.

https://github.com/matterport/Mask_RCNN/releases/tag/v2.0

These are the results on the COCO 5K minival. Bounding box AP is 34.7 (the paper reports 38.2 on test-dev but they don't report minival, unless I missed it). For segmentation, the AP is 29.6 (the paper reports 35.4 on minival). Playing with hyper parameters might allow you to get higher accuracy, but I didn't have enough time to try.

Evaluate annotation type *bbox*
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.347
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.544
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.377
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.163
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.390
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.486
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.295
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.424
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.433
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.214
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.481
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.601

Evaluate annotation type *segm*
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.296
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.510
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.306
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.128
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.330
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.430
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.258
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.369
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.376
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.173
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.417
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.538

Hi, Do you have any notes on the hyper parameters, data augmentation, optimizer etc.?

I am working on the code as is with no changes made and am trying to train on Coco2014 dataset
"bbox" results:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.142
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.235
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.147
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.048
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.166
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.248
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.157
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.192
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.193
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.051
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.192
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.354
Prediction time: 173.85491013526917. Average 0.3477098202705383/image
Total time: 182.46349167823792

I have read through many issues here and have picked up various suggestions - use Adam, data augmentations, use ResNet50, lower learning rate, train all layer directly without stage-wise training - to mention a few.

I have a couple of specific questions regarding training on Coco dataset.

How many batches (model sees the whole dataset once) do you train for?
What is the best train/val (absolute) loss values? (so I know what to expect from the train logs)
The general train loss values i see at my end is something like:
loss: 1.8633 - rpn_class_loss: 0.0268 - rpn_bbox_loss: 0.7644 - mrcnn_class_loss: 0.3761 - mrcnn_bbox_loss: 0.4218 - mrcnn_mask_loss: 0.2742
where in only rpn_class_loss is small. Other loss values are relatively larger (as seen here)

Finally, generally, I am wondering if you have some suggestions from your experiments on the right direction to get the model to a convergence point?

ashwin-999 · 2021-06-22T22:11:12Z

@stark-xf Hi, could you figure how to the model trained?

waleedka mentioned this issue Nov 26, 2017

Questions on Masks and class predictions #88

Closed

fastlater mentioned this issue Mar 8, 2018

how to convert your dataset into annoted format. #155

Open

lasseha mentioned this issue Mar 16, 2018

KeyError: 'unexpected key "conv1.weight" in state_dict' multimodallearning/pytorch-mask-rcnn#3

Closed

Cpruce mentioned this issue Apr 11, 2018

Mobile Mask RCNN #306

Open

ppwwyyxx mentioned this issue Jul 5, 2018

Is there ResNet-101 pretrained imagenet weights.h5 #725

Open

ppwwyyxx mentioned this issue Feb 18, 2019

Advantages compared to tensor-flow version Mask-RCNN facebookresearch/maskrcnn-benchmark#449

Open

ppwwyyxx mentioned this issue Apr 6, 2019

Can this implementation achieve the state-of-art performance? #1412

Open

This was referenced May 14, 2019

Inference performance with different backbone #1420

Open

When i use the mobilenet as a backbone,i met this mistake? #558

Closed

Spiffy1 mentioned this issue Sep 20, 2019

Detection Returning Empty & Model Containing Dead Neurons #753

Open

ppwwyyxx mentioned this issue Apr 12, 2020

Reproducing the results of coco dataset #2097

Open

ajaysub110 pushed a commit to ajaysub110/Mask_RCNN that referenced this issue Sep 13, 2020

Merge pull request matterport#1 from akTwelve/log_dir_fix

7aed1f9

Bugfix for MaskRCNN creating empty log_dir that breaks find_last()

elivety mentioned this issue Sep 17, 2020

Multi class Instance segmention on custom data #2348

Open

NicholasCF pushed a commit to NicholasCF/Mask_RCNN that referenced this issue Feb 10, 2021

Merge pull request matterport#1 from leekunhee/tensorflow2.0

111ab9c

Tensorflow2.0

UsmanAfzaal mentioned this issue Jun 1, 2021

MSCOCO evaluation results are "Zero". #2582

Open

deveshdatwani mentioned this issue Dec 11, 2021

Correct format for polygons? #2741

Closed

Do you report MSCOCO performance? #1

Do you report MSCOCO performance? #1

Comments

gaopeng-eugene commented Nov 1, 2017

waleedka commented Nov 2, 2017

larsoncs commented Nov 8, 2017 • edited Loading

yanxp commented Nov 9, 2017 • edited Loading

ppwwyyxx commented Nov 9, 2017

vijaykbg commented Nov 10, 2017

waleedka commented Nov 10, 2017

larsoncs commented Nov 10, 2017

vijaykbg commented Nov 10, 2017

csnemes2 commented Nov 13, 2017

csnemes2 commented Nov 13, 2017

Walid-Ahmed commented Nov 13, 2017

ypflll commented Nov 17, 2017

ypflll commented Nov 17, 2017

ypflll commented Nov 17, 2017

jiang1st commented Nov 20, 2017 • edited Loading

ppwwyyxx commented Nov 20, 2017 • edited Loading

jiang1st commented Nov 21, 2017

ypflll commented Nov 21, 2017

csnemes2 commented Nov 21, 2017

waleedka commented Nov 26, 2017

ppwwyyxx commented Nov 26, 2017 • edited Loading

jiang1st commented Nov 26, 2017 • edited Loading

waleedka commented Nov 27, 2017

waleedka commented Nov 27, 2017

ppwwyyxx commented Nov 27, 2017

waleedka commented Nov 27, 2017

jmtatsch commented Nov 27, 2017

jmtatsch commented Dec 10, 2017

jiang1st commented Dec 11, 2017

John1231983 commented Feb 2, 2018

John1231983 commented Feb 2, 2018

John1231983 commented Feb 2, 2018

John1231983 commented Feb 3, 2018

jmtatsch commented Feb 5, 2018

endluo commented Jul 18, 2018 • edited Loading

michaelisc commented Aug 6, 2018

s-bayer commented Aug 8, 2018

s-bayer commented Aug 13, 2018

michaelisc commented Aug 13, 2018

s-bayer commented Aug 14, 2018

stark-xf commented Oct 17, 2018

SG-97 commented Sep 7, 2020

ashwin-999 commented Jun 22, 2021

ashwin-999 commented Jun 22, 2021

larsoncs commented Nov 8, 2017 •

edited

Loading

yanxp commented Nov 9, 2017 •

edited

Loading

jiang1st commented Nov 20, 2017 •

edited

Loading

ppwwyyxx commented Nov 20, 2017 •

edited

Loading

ppwwyyxx commented Nov 26, 2017 •

edited

Loading

jiang1st commented Nov 26, 2017 •

edited

Loading

endluo commented Jul 18, 2018 •

edited

Loading