Hello!
Problems of CLUSTERS.

Is CLUSTERS=3 used in Yet-Another-EfficientDet-Pytorch?

Now I want to use your code to get the anchor of my own dataset, but I found that you added '' category_id='l' '' when calculating the anchor. It seems that the purpose is to calculate merely the anchors of large-scale target.

So, what should I do to get the right anchors. Should I set 'category_id' to a string other than 's, m, l'?

I came across this repo while browsing issues in zylo117's EffDet. Great tool btw. Thanks for hosting.

Zylo's effdet implementation uses format (w,h) for his anchor box implementation. When I pass a set of ratios that have boxes which are wider and shorter (That's how the gt boxes are too), according to zylo's format, they'd be something like:
anchors_ratios: '[(0.9, 1.1), (1.2, 0.8), (2.2, 0.5)]'

When I pass it to the anchor inspector along with the training data directory, I see that I see a log of red boxes. Some green.
However, If i pass the inverted ratios, lots of green start popping up:
eg: [(1.1,0.9), (0.8,1.2), (0.5,2.2)]

So, is it possible that the format required for the anchor inspector is (h,w)?

I tried to re-calculate anchor_scale and anchor_ratio on my own dataset using method in your repo. The result accuracy is 58.11% for CLUSTER=3, what do you think about this result? Is that satisfactory？

Also, what does the parameter CLUSTER mean? I notice it controls the length of output list anchor_scale and anchor_ratio. I use EfficientDet from "Yet-Another-EfficientDet-Pytorch" and the length anchor_ratio is 3 in that repo, so I should set CLUSTER=3 in this case?
Thanks!

hello, thanks for your work for this repo.
I wonder how to set the params like anchor_base_scale & anchor_stride for efficientdet?
Could you tell me the way to get above two ?

Hi there, please could you provide a link to the dataset in the notebook example so I may inspect the xml files as I have some errors I would like to debug. Thank you.

Thank you for your great work! But I want to konw how to get customer anchor in Yet-Another-EfficientDet-Pytorch.

Hi @Cli98, I think that in the current implementation the bbox size has too much importance.

Immagine this situation:
10 bboxes 10x5
5 bboxes 2x1
5 bboxes 1x2

If K=2 the two clusters would be:
10 bboxes 10x5

and
5 bboxes 2x1
5 bboxes 1x2

and so the suggested ratios would be (1.4, 0.7) and (1., 1.) instead of (1.4, 0.7) and (0.7, 1.4) (or (1., 0.5), (1., 2.)).

I tried this experiment with COCO annotations, I'm trying to calculate good anchors for EfficientDet0 and my dataset:

# load COCO train 2017 annotations
with open("instances_train2017.json") as f:
    annotations = json.load(f)

image_size = 512 #efficientdet0 input_size
min_size = 32 * 32
images_scale = {ann["id"]: image_size / max(ann["width"], ann["height"]) for ann in annotations["images"]}
scaled_bboxes = np.array([np.array(ann["bbox"][-2:]) * images_scale[ann["image_id"]] for ann in annotations["annotations"] if np.prod(ann["bbox"][-2:]) > min_size])
out = kmeans(scaled_bboxes, k=3)
# results:
2 * out / out.sum(axis=1, keepdims=True)
array([[0.91217741, 1.08782259],
       [0.914852  , 1.085148  ],
       [1.00598999, 0.99401001]])

As you see the ratios are very different from the suggested ones and very similar to each other.

Instead if I normalize the bboxes so they sum to 2 (like in EfficientDets libraries):

min_size = 32 * 32
bboxes = np.array([annotation["bbox"][-2:] for annotation in annotations["annotations"] if np.prod(annotation["bbox"][-2:]) > min_size])
nbboxes = 2 * bboxes / bboxes.sum(axis=1, keepdims=True)
out = kmeans(nbboxes , k=3)
# results:
2 * out / out.sum(axis=1, keepdims=True)
array([[0.60458624, 1.39541376],
       [0.96894304, 1.03105696],
       [1.33333333, 0.66666667]])

Here as you can see we obtain more or less the ratios suggested, for example, in the EfficientDet implementation

I enabled --anchors and --annotations,

but no annotations shows and no anchors shows (using defrault params)

my dataset is:(according to Yet-Another-EfficientDet-Pytorch)

dataset
object
annotations
train
val

And I think all the settings are right

when running program：
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
Loading image with numpy array
Image id: 0
Loading image with numpy array
Image id: 1
Loading image with numpy array
Image id: 2
Loading image with numpy array
Image id: 3

No task results，What's the problem？？？？？？？？？？

Hi, Bro. It haunts me much. The question is as followed:
I am using the images of my uav. And the image size is 38402160, the object i want to detect is 94325 or 236938, so is the object small or mid or large?(According to the coco, it is absolutely large cause it is bigger than 9696, but how am i going to process the 3840*2160 images when put it into yolo series or ssd?)
I will appreciate it if can i get your reply soon. Thanks.

I was wondering if you can give me a template yaml files, and template xml annotation file.
the below folder structure is within anchor_computation_tool.

+---projects
| mask.yml
+---datasets
| +---mask
| | | instances_train.json
| | | instances_val.json
| +---train
| | ---images
| | train1.png
| | train10.png
+---val
| | ---images
| val5.png
| val6.png

the instances_train.json are COCO annotation that have bounding box and annotation.
The mask.yaml has the project according to ZYLO117 recommendation.

I need help on how to format the xml annotation file from the json file.
Thanks,

there were no such package?

Hi @Cli98 ,

Could you share me code for json file, please?
Thank you so much