Darknet: Trailing comma not removed from coco_results.json in Windows

Created on 8 Jun 2019 · 2Comments · Source: AlexeyAB/darknet

There appears to be an issue caused by the difference between Windows vs Linux line-endings, when running a COCO evaluation. When running the detector on the COCO dataset, e.g. "detector valid cfg/coco.data ../cfg/yolov3.cfg ../weights/yolov3.weights", a "results/coco_results.json" file is generated. After running on a Linux system, an example of the last few lines of this file is:

{"image_id":284282, "category_id":65, "bbox":[0.000000, 18.000229, 106.680382, 308.656616], "score":0.010804},
{"image_id":284282, "category_id":47, "bbox":[350.594543, 30.112206, 36.254150, 39.351761], "score":0.112114}
]

However, when run under Windows, the output includes a trailing comma after the last entry, e.g.:

{"image_id":284282, "category_id":65, "bbox":[0.000000, 18.000229, 106.680382, 308.656616], "score":0.010804},
{"image_id":284282, "category_id":47, "bbox":[350.594543, 30.112206, 36.254150, 39.351761], "score":0.112114},
]

Because a trailing comma is typically not allowed in a JSON object, the subsequent COCO evaluation python scripts fail when parsing a coco_results.json file generated from a darknet run on Windows.

Currently, detector.c (lines 557-559) and coco.c (lines 221-223) are:

 fseek(fp, -2, SEEK_CUR);
 fprintf(fp, "\n]\n");
 fclose(fp);

After running darknet, an ASCII dump of the end of coco_results.json from Linux is:

162447420   9   .   3   5   1   7   6   1   ]   ,  sp   "   s   c   o   r
162447440   e   "   :   0   .   1   1   2   1   1   4   }  nl   ]  nl

From Windows, it's:

163477420  sp   "   s   c   o   r   e   "   :   0   .   1   3   1   5   5
163477440   7   }   ,  cr  nl   ]  cr  nl

So for coco evaluation to work properly on Windows, it looks like the fseek() offset should be -3 on Windows and -2 on Linux. Another solution may be to pass a flag to print_cocos() so that the comma delimiter is not printed for the last line.

Bug fixed

Source