You Only Look Once ( YOLO v1 )

Implementation of YOLO object detection pipeline using tensorflow. YOLO is a real time object detection method. It treats both object detection and localisation as regression problems. This is in contrast to previous object detection pipelines such as R-CNN, which had seperate entities for detection and localisation and were far more complicated to fine tune/train. More on YOLO here.

As of now, YOLO v2 is out. Check it out here.

How to use ?

Make sure that weight file is present in weights directory. Currently there are three modes, all pertaining to test the pre-trained model.

• 'testDB' - Tests the code on a database. ( PASCAL VOC 2007, 2012, MS-COCO ). Keep in mind the model has been trained on PASCAL VOC 2007+2012. So any other dataset would require training. As this has not been implemented yet, darkflow may help.
• 'testLive' - Tests from a live Webcam feed.
• 'testFile' - Tests on a single image.

By default, it runs on 'testLive' mode.

python yolo.py

Results (PACAL VOC 2007)

Class Name	Ground Truth	Predicted	True Positive	False Positive	Avg. Precision
aeroplane	311	213	141	72	0.55494075262829001
bicycle	389	237	157	80	0.56105639251746608
bird	576	359	184	175	0.42433899865958929
boat	393	213	77	136	0.24679748475368918
bottle	657	128	33	95	0.17272727272727273
bus	254	168	117	51	0.54621080695222668
car	1541	925	436	489	0.34186953795331443
cat	370	322	250	72	0.67658801636799115
chair	1374	420	102	318	0.12245608573113981
cow	329	204	66	138	0.17318304265255621
diningtable	299	160	114	46	0.6494860956834515
dog	530	422	299	123	0.63538308205967287
horse	395	279	209	70	0.63316214093397671
motorbike	369	228	140	88	0.48458388143892261
person	5227	3319	1166	2153	0.23070649513423477
pottedplant	592	200	53	147	0.17236723672367235
sheep	311	172	44	128	0.16292819499341238
sofa	396	141	105	36	0.6494177280693908
train	302	255	191	64	0.66130285346624584
tvmonitor	361	209	133	76	0.54199124564843304

Requirements

• Tenseflow 1.0
• OpenCV 2
• Python 2
• Pre-trained weights

TODO List

• Complete the mean-Average Precision
• Document the code
• Add PASCAL VOC 2007 results to the readme
• Complete network training function

References

• Author's Website
• This implementation