Reproducibility of ActivityNet about tadtr HOT 4 CLOSED

Hi, thanks for your interest in our work. To reproduce our results with TSP features on ActivityNet, some parameters need to be modified. For example, num_queries should be 50, set_cost_class should be 2. Besides, several tricks (such as soft-NMS, using external video classification labels from CUHK's winning model at ActivityNet 2017) from BSN/BMN are required to boost performance.
I do plan to release the code for ActivityNet. But it might be one or two months later. If you need it recently, you can write an email to me for code request.

from tadtr.

Really thanks for your reply. This helps me a lot. I have sent you a email for code request.

from tadtr.

Is Soft-NMS beneficial for this query-based method?
How does it work?

from tadtr.

Hi @takfate, here is my response to your questions:
Soft-NMS is beneficial on ActivityNet but harmful on THUMOS14, and HACS, according to my experience. The code already includes the implementation of Soft-NMS. To enable that, you need to:

1. change the argument nms_mode from ['raw'] to ['raw', 'nms'] in line 115 of engine.py:
2. change line 151~156 of datasets/tad_eval.py to

if nms_mode == 'nms' and not (config.TEST_SLICE_OVERLAP > 0 and self.dataset_name == 'thumos14'):  
    # On THUMOS14, when config.TEST_SLICE_OVERLAP > 0, 
    # we only apply nms after all predictions have been collected
    dets = apply_nms(input_dets,
                     nms_thr=config.NMS_THR,
                     use_soft_nms=self.dataset_name in ['activitynet'])
else:
    sort_idx = input_dets[:, 2].argsort()[::-1]
    dets = input_dets[sort_idx, :]

from tadtr.