MaskDino Fails to learn Precise Bounding Boxes on custom dataset but Dino does

[FabianSchuetze]

Thanks for the wonderful repo. It's a pleasure to work with it and to read the code.

When training MaskDino on a custom dataset, the bounding box predictions are not very good. Interestingly:

• Dino learns good bounding boxes on the same dataset
• The instance masks predicted by Mask Dino are good too
• Bounding boxes with 50mAP are learned with MaskDino, but the bounding boxes for mAP 75 (and better) are poor.
• MaskDino errored at the end with a problem in the cost_matrix. See the logs for details.

Does anybody have an idea what I could tune to generate good bb results?

Training Details:
I have slightly modified the training process (see this branch https://github.com/FabianSchuetze/detrex/tree/my_changes). I added amp training and have included some gradient checkpointing. I train with one GPU and a batch size of four (for MaskDino, Dino works with a batch size of 8). The learning rate is lowered linearly.

Data:
The instances are very dense, similar to the "is-crowded" scenes of COCO. There is only one class. I have adjusted the num_objects in the config files.

Logs:
Logs of the training runs are attached below. There are three logs:

• One for MaskDino with the original noise scale (0.4)
• One with a noise scale of 1.0 (Dino uses this value)
• One for Dino (noise scale of 1.0)

Hyparameters:
Comparing the parameters, the following aspects seem notable:

• Maskidio has a higher Hungarian class loss (5 vs 2)
• Maskdino has 1/3 of the queries (300 vs 900)

maskdino_0.4_noise_scale.txt
maskdino_1.0_noise_scale.txt
dino_log.txt

Does anybody have an idea how to debug the problem?

[FabianSchuetze]

To reproduce the results, I have used a public dataset with similar characteristics. In The COB-3D dataset, see: https://arxiv.org/abs/2210.07424 . I have extracted rgb images, bounding boxes, instance mask in the coco format. The dataset is a bit small (~6k images) and can be downloaded here. The original data is here. Please note that the data is published under the CC, non-comercial see https://github.com/wyndwarrior/autoregressive-bbox/blob/main/LICENSE .

An image of the predictions with maskdino and the gt are:

The logs for dino and mask dino are uploaded below.
dino.txt
maskDino.txt

Interestingly:

• The bb mAP for dino is much better than for maskDino. The training is a bit short, but I noticed similar difference after longer training
• However, when looking at the predictions, the visualized bbs for dino at not that much better. Both show a little bit of a low recall. I also uploaded the json predictions.
• The results with standard Mask-RCNN heads are generally pretty good on this dataset. They have a good recall and good precision.

[HaoZhang534]

To reproduce the results, I have used a public dataset with similar characteristics. In The COB-3D dataset, see: https://arxiv.org/abs/2210.07424 . I have extracted rgb images, bounding boxes, instance mask in the coco format. The dataset is a bit small (~6k images) and can be downloaded here. The original data is here. Please note that the data is published under the CC, non-comercial see https://github.com/wyndwarrior/autoregressive-bbox/blob/main/LICENSE .

An image of the predictions with maskdino and the gt are:

The logs for dino and mask dino are uploaded below. dino.txt maskDino.txt

Interestingly:

• The bb mAP for dino is much better than for maskDino. The training is a bit short, but I noticed similar difference after longer training
• However, when looking at the predictions, the visualized bbs for dino at not that much better. Both show a little bit of a low recall. I also uploaded the json predictions.
• The results with standard Mask-RCNN heads are generally pretty good on this dataset. They have a good recall and good precision.

Hello, I notice that the boxes by maskdino are all shifted upper right a little bit. I guess there may be some bugs in the postprocessing code.

[HaoZhang534]

@FabianSchuetze When you have relatively small datasets, Mask-RCNN usually can do good enough. MaskDINO and DINO are suitable for relatively large datasets such as COCO.

[HaoZhang534]

@FabianSchuetze We fixed a bug in #249. Maybe you can run again to see if this solved your problem. Please also refer to the discussions in #247 .

[FabianSchuetze]

Thank you so much, @HaoZhang534 ! I will train the model again tomorrow and report back.

[FabianSchuetze]

@HaoZhang534 . I have worked wit the new commits but the bounding boxes are still shifted. I have commented again in #247 .

Furthermore, I am still not getting very good results. Maybe the training process is not really possible with just a batch size of 4? I will try to train on MS CoCo and see whether I can reproduce the original results. Can you maybe attach a log of the original training process? That would be wonderful & would make a comparison easier.