Divyam Anshumaan

icon location Research Fellow at Microsoft Research. Previously at IIITD. Usually working on something interesting.

Results 11 comments of Divyam Anshumaan

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

Did you scale your annotations with the images?

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

I was trying to train it on SIM10k -> Cityscapes myself. Could you help here? We can compare our results afterwards. Traceback: before filtering, there are 5950 images... after filtering,...

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

I tested the domain adaptation model. The car AP is 40.1 with a baseline of 30.6. The only thing I did differently was to keep all the pascal voc labels...

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

Source data was SIM10K. The only class in pascal_voc.py should be 'car', but for convenience I took a super set which included the classes from the pascal voc dataset and...

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

I don't think it should be a problem but let me check. As for the different number of images, I kept a try except statement that would initialize the target...

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

[modified_files.zip](https://github.com/tiancity-NJU/da-faster-rcnn-PyTorch/files/3268573/modified_files.zip) Here you go!

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

Also, test ouput: Called with args: Namespace(cfg_file='cfgs/vgg16.yml', checkepoch=4, checkpoint=19999, checksession=101, class_agnostic=False, cuda=True, dataset='city', large_scale=False, load_dir='./data/pretrained_model', mGPUs=False, model_dir='models.pth', model_name='res101.bs1.pth', net='vgg16', parallel_type=0, set_cfgs=None, vis=False) Using config: {'ANCHOR_RATIOS': [0.5, 1, 2], 'ANCHOR_SCALES': [4,...

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

@xiong233 The baseline mAP that I have from my test was something around 29. Will run the test again and let you know.

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

trained on sim10k, tested on cityscapes with VGG16. Called with args: Namespace(cfg_file='cfgs/vgg16.yml', checkepoch=3, checkpoint=19999, checksession=517, class_agnostic=False, cuda=True, dataset='city', large_scale=False, load_dir='./data/pretrained_model/', mGPUs=False, net='vgg16', parallel_type=0, set_cfgs=None, vis=False) Using config: {'ANCHOR_RATIOS': [0.5, 1,...

I intend to reproduce SIM10k to cityscape, but car AP's AP is getting lower and lower as the number of training iterations increases.

[standard_train.zip](https://github.com/tiancity-NJU/da-faster-rcnn-PyTorch/files/3293100/standard_train.zip) [modified_files.zip](https://github.com/tiancity-NJU/da-faster-rcnn-PyTorch/files/3293098/modified_files.zip) Here you go. standard_train.zip has files for standard training. I used this for the above model. Please let me know if there are any mistakes on my end....