Open tjpreddy opened 3 years ago
Did you solve this problem I have the same issue
I don't think they fully implement the code correctly, as in the original paper the ball detection module has two last fully connected layers one produces X and one produces Y representing the coords. However, in their code, they have only one that produces a vector in the shape of width + height = 448. So there is some mismatch which can be the possible reason why the ball detection stage in this repo is not working so well when comparing to original paper.
Just want to add, you can try adding --thresh_ball_pos_mask 0.00001
to the test.sh file which should allow it to produce valid output.
Trained TTNet global model using ttnet_1st_phase.sh. Using the above trained model for new test sequence always gives fixed prediction for the entire video. python test.py --working-dir ../ --saved_fn ttnet_1st_phas --no-val --batch_size 8 --num_workers 1 --lr 0.001 --lr_type 'step_lr' --lr_step_size 10 --lr_factor 0.1 --global_weight 5. --seg_weight 1. --no_local --no_event --no_seg --smooth-labelling --show_image here!! number of trained parameters of the model: 7781952 loading pre-trained model 0%| | 0/54 [00:00<?, ?it/s] ===================== batch_idx: 0 ================================
Ball Detection - Global stage: (x, y) - gt = (199, 57), prediction = (0, 72) Ball Detection - Overall: (x, y) - org: (1196, 486), prediction = (0, 607)