Conversation
|
@alpha-carinae29 thanks a lot for this PR. Do you have any quantitative or qualitative results to compare performance of the detection before and after applying these filters? |
Yeah I have computed the mAP for last 1000 frames of Oxford Town Center dataset, however since the models are trained on this environment and have high accuracy and also there are relatively few number of frames for computing the prior distributions, there was not a distinguishable improvement in accuracy. We had just 0.2% improvement in mAP. |
|
@alpha-carinae29 can you please resubmit this PR to the smart-social-distancing repo? |
| [PostProcessor] | ||
| MaxTrackFrame: 5 | ||
| NMSThreshold: 0.98 | ||
| CalibrationFrames: 100 |
There was a problem hiding this comment.
please provide explanation about post processing config parameters.
| 3. apply a simple object tracker to make the detection more robust. | ||
| 4. filter bounding boxes based on boxes prior distribution. | ||
| 5. filter bounding boxes based on the amount of foreground pixels exist on the box. | ||
|
|
There was a problem hiding this comment.
- Post processing step is getting complex and it's good to have a detailed documentation outside of code.
- I have an idea to develop a scoring algorithm for each bounding boxes based on these 5 steps criteria and detection score to omit extra bounding boxes in a unified way. each of steps add or reduce each bounding box acceptance score with a weighted score and bounding boxes with a score greater than a threshold will remain.
|
@alpha-carinae29 any updates on this? |
|
@mhejrati, can you review this? |
|
migrated to: |
3 participants
I implemented two box filtering strategy based on background subtraction and box prior distributions.
this will close #30