amoussawi / caffe Goto Github PK

Hi @amoussawi , I have two questions in my SSD training on Caltech.

1. Have you used hard negative mining during training?

2. In your paper, you have mentioned the ETH and TUD-Brussels dataset results additional 4000 images, but I could only find 2313 images with pedestrian annotations.
   (ETH setup1 ~ set00 V000 999
   set01 V001 451
   set02 V002 354,
   TUD ~ 507)

Hi Abdallah,
Thank you for the great work and sharing your code.

I have few questions about training settings. How did you know that after 20K iterations the model starts to overfit ? and how many training images did you have, and the number of epochs this 20K represents ?

Thank you.

Hi
Thx for your excellent excellent excellent job! I'm a new learner and it benefit me a lot.

My question is: i wonder which of these two aspect_ratios configuration work better. How much better ? and How about the distinction of time cost between them?

(1) in ssd_caltech_512.py
331 aspect_ratio_per_layer = [0.15,0.3,0.41,0.6,0.75,0.9]
332 aspect_ratios = [aspect_ratio_per_layer]*len(mbox_source_layers)
(2) in ssd_caltech_512_ft.py
334 aspect_ratio_per_layer = [0.41]
335 aspect_ratios = [aspect_ratio_per_layer]*len(mbox_source_layers)

Another question is : i can only run the SSD512 model with the speed about 12 FPS (Geforce GTX Titan X), how can i achieve 24 FPS like u?

Thx again again again :) @amoussawi

Hi @amoussawi
Thanks for sharing the code. I'm trying to apply SSD on Caltech too, so I have a few questions here:
1、How do you deal with annotations like 'person?' or 'people' in your trainval set? Also, did you use all ground truth annotations for training, regardless their heights and occlusions?
2、Since the detection output and evaluation code is for VOC, how do you evaluate your results in Caltech style?
3、Have you tried training only with caltech trainset? I'm interested in how much the performance will boost using extra data(ETH,...) for training.

Thanks for your sharing your excellent work. Did you select the pedestrians greater than 30 or 50 pixels to train your code? Thanks

Hi,

First of all, thank you for sharing your code!

I have a question/issue: I'm trying to run the provided pre-trained model with the original SSD code, but I'm getting a couple of errors. First, the field "extra_ar" in the PriorBox layers of the deploy.prototxt cannot be read: Error parsing text-format caffe.NetParameter: 1010:13: Message type "caffe.PriorBoxParameter" has no field named "extra_ar"

If such fields are removed, then another error appears: F0627 07:17:17.537052 12113 detection_output_layer.cpp:164] Check failed: num_priors_ * num_loc_classes_ * 4 == bottom[0]->channels() (163836 vs. 54612) Number of priors must match number of location predictions.

For what I've seen, your have modified certain things in the SSD architecture, such as the aspect ratio in the layers. However, I though such modifications were self-contained, i.e. they only affected definitions and scripts, but that the resulting deploy.prototxt and .caffemodel were compatible with mainstream SSD. Is that the case, or is your branch needed to run the model for inference? Any tips or explanations about the error are welcomed!

Thanks!

Issue summary

Hi, first I want to thank you for the effort of doing Caltech on SSD. Here is my question:
I do step 4 in the Preparation step(I didn't do step 3 since I don't need evaluation yet), and I put the downloaded Caltech file in /home/data/caltech_code/. But when I execute ./data/caltech/creat_list.sh, it shows these two messages:

Create list for caltech trainval...
ls: cannot access '/home/paris/caffe/data/caltech/trainval/images//': No such file or directory
ls: cannot access '/home/paris/caffe/data/caltech/trainval/images//': No such file or directory
and ...
Create list for caltech test...
ls: cannot access '/home/paris/caffe/data/caltech/test/images//': No such file or directory
ls: cannot access '/home/paris/caffe/data/caltech/test/images//': No such file or directory

I'm don't know where I've done wrong or do I miss any step?

Hi, very glad to see your excellent work. I'd like to know where I can get the full text of your essay:An FPGA-Accelerated Design for Deep Learning Pedestrian Detection in Self-Driving Vehicles. Please help. Thx.

@amoussawi Hi, thanks for your work. You mentioned that the miss rate is 11.8% after 20k iterations.

1. And I'd like to know what is the detection_eval that the network gives out on test set after 20k iterations?
2. I get the detecion_eval = 0.545697 after 8k iterations with fine-tuning. I think it's a little bit low. Is it normal or maybe something is wrong? What's your opinion?
   Thanks.