Low mAP on Activitynet-1.3

[yangmin666]

When I run all the codes to obtain detection_result_nms0.6.json, the detection results is bad.

[yangmin666]

Another question: What is "gsm_best_score.json" in spot_inference.py? Your repository does not contain it. Thank you.

[sauradip]

Can you paste the config file here ? Otherwise it is difficult for me to judge ! And also did you maintain the same environment as mentioned in the Readme ? I can reproduce the results by training.

[sauradip]

"gsm_best_score.json" should be renamed to "spot_best_score.json". I will correct it. Thanks for spotting

[yangmin666]

Thank you! I fix this error now. It is 29.912 mAP.

[yangmin666]

"gsm_best_score.json" should be renamed to "spot_best_score.json". I will correct it. Thanks for spotting

However, I think using "spot_best_score.json" in semi-supervised method is incorrect. When I set full_label to False, the mAP drops.

[sauradip]

We follow the same technique as SSTAP. The post-processing in most of the TAD approaches irrespective of classifier is still the same

[zhangquan97]

Thank you! I fix this error now. It is 29.912 mAP.
Hello, my reproduction performance is the same as yours when using 10% annotated data, but the mAP of 29.912 is much lower than the mAP of 32.1 in the paper. Do you know the reason for this?

[zhangquan97]

Besides, I am concerned about the getAnnotation() function used multiple times in spot_dataloader. py is very confusing

                if action_end - action_start > 1:
                    label_list.append([snip_start,snip_end,gt_label])  

Should these two lines be placed inside the loop for j in range (len (labels)):?

When I placed these two rows outside the loop, I reproduced the performance of mAP34.13 using 10% annotated data. When I placed them outside, the performance reproduced was 29.912.