Giters

hosang / detection-proposals

Evaluation of detection performance, recall, and repeatability of detection proposal algorithms

Home Page:http://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-computing/research/object-recognition-and-scene-understanding/how-good-are-detection-proposals-really/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Unclear if detection proposals required to train at all or just.. train more quickly

EMCP opened this issue · comments

commented

This question is more general, so apologies if it's not quite about this repo.. but I am wanting to train faster R-CNN on a custom COCO formatted dataset.

I am using the project saved here : https://github.com/jwyang/faster-rcnn.pytorch

In it, they explain to download COCO2014 ... and eventually it comes to a link to your proposals. My question is, are the proposals necessary to train? It seems that the answer should be no, but

If I want to more quickly come to a good object detection model on my custom dataset .. I had best create my OWN object proposals... ? Do I try all of them? it's unclear how I run them all.. but I imagine it's somewhere in this https://github.com/hosang/detection-proposals#format-of-the-precomputed-proposals ?

so in short, my goal is a custom precomputed-coco/ .. ideally with the following sub-folders

drwxr-xr-x  3 emcp emcp 4096 mar 10 20:32 BING
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:44 edge_boxes_70
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:48 edge_boxes_AR
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:51 endres
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:52 geodesic
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:40 MCG
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:53 random_gaussian
drwxr-xr-x  3 emcp emcp 4096 mar 10 20:59 randomized_prims
drwxr-xr-x  3 emcp emcp 4096 mar 10 21:02 rigor
drwxr-xr-x  3 emcp emcp 4096 mar 10 21:03 segmentation_baseline
drwxr-xr-x  3 emcp emcp 4096 mar 10 21:04 selective_search
drwxr-xr-x  3 emcp emcp 4096 mar 10 21:13 sliding_window

I do not have access to matlab, so it sounds like I would need to sort out implementations of these object detection proposal techniques / networks elsewhere, correct?

commented

I think I get it now... based on this article

https://medium.com/@fractaldle/guide-to-build-faster-rcnn-in-pytorch-95b10c273439

indeed FASTER RCNN is all about building networks themselves that can predict objects.. and now I am wondering why my example project is pulling in your coco-precomputed/ .. perhaps to train?