• Hardware Platform GPU
• DeepStream Version 6.3
• TensorRT Version 8.6.1.6
• NVIDIA GPU Driver Version 510.73.08
• How to reproduce the issue ?
I’m deploying a classification model, but I’m encountering an issue where frame_meta.obj_meta_list returns None, which prevents me from obtaining the resulting class. Does anyone know what might be wrong or how to resolve this problem?
if (info.type & Gst.PadProbeType.BUFFER):
gst_buffer = info.get_buffer()
if not gst_buffer:
logger.warning("Unable to get GstBuffer")
return
buffer_received_time = datetime.now().strftime('%H:%M:%S.%f')
batch_meta = pyds.gst_buffer_get_nvds_batch_meta(hash(gst_buffer))
l_frame = batch_meta.frame_meta_list
cont = 0
# Loop para percorrer cada frame
while l_frame is not None:
time_now = datetime.now()
time.sleep(0.1)#Tempo de espera entre cada frame, estou usando para analisar as saidas
try:
frame_meta = pyds.NvDsFrameMeta.cast(l_frame.data)
#cls_meta=pyds.NvDsClassifierMeta.cast(frame_meta.data)
#cls_meta_lbl = cls_meta.label_info_list
#classifier = frame_meta.base_meta.batch_meta.classifier_meta_pool
#print('frame_meta:',classifier)
# label_info_list = pyds.NvDsLabelInfo.cast(classifier.label_info_list.data)
#print('label_info_list.result_class_id: {}'.format(label_info_list.result_class_id ))
#print('cls_meta.num_labels:',cls_meta.num_labels)
# print('cls_meta_lbl')
#cls_meta_lbl_info=pyds.NvDsLabelInfo.cast(cls_meta_lbl.data)
#result_str = str(cls_meta_lbl_info.result_label)#.tobytes().decode('iso-8859-1'))
#print("result_decode:", result_str)
except Exception as e: #StopIteration
break
#print(help(frame_meta))
#No primeiro frame pega altura e largura e atualiza o tempo da ultima deteccao para nao entrar em stop line
if self.frame_width == 0:
self.frame_width = frame_meta.source_frame_width
self.frame_height = frame_meta.source_frame_height
self.last_detection_moviment_timestamp = datetime.now()
self.time_continuous_detection = 0
print(frame_meta.source_frame_width, frame_meta.source_frame_height)
#Display meta para adicionar textos na imagem
display_meta=pyds.nvds_acquire_display_meta_from_pool(batch_meta)
# Me remova
self.cont_frame += 1
#Caso esteja em parada de linha, adicionar mensagem na imagem
if self.stop_line_status:
#Mensagem na imagem
self.add_text_to_display_meta(display_meta, "PARADA DE LINHA", int(self.frame_width/2), int(self.frame_height/2),1,1, font_size=40)
else:
#Adicionando no frame linha central que serve para contagem, isso é só para visualização
self.add_lines_to_display_meta(display_meta, self.line_count_x, 0, self.line_count_x, self.line_count_y)
#Objeto que contem as deteccoes
l_obj = frame_meta.obj_meta_list
#print(l_obj.class_id)
print("l_obj:",l_obj)
print("frame_meta.num_obj_meta:",frame_meta.num_obj_meta)
print("frame_meta.bInferDone:",frame_meta.bInferDone)
Thank you for answer. Since l_obj is None, it’s not possible to retrieve obj_meta, which means accessing its classifier_meta_list method is also not feasible.
if (info.type & Gst.PadProbeType.BUFFER):
gst_buffer = info.get_buffer()
if not gst_buffer:
logger.warning("Unable to get GstBuffer")
return
buffer_received_time = datetime.now().strftime('%H:%M:%S.%f')
batch_meta = pyds.gst_buffer_get_nvds_batch_meta(hash(gst_buffer))
l_frame = batch_meta.frame_meta_list
cont = 0
# Loop para percorrer cada frame
while l_frame is not None:
time_now = datetime.now()
time.sleep(0.1)#Tempo de espera entre cada frame, estou usando para analisar as saidas
try:
frame_meta = pyds.NvDsFrameMeta.cast(l_frame.data)
#cls_meta=pyds.NvDsClassifierMeta.cast(frame_meta.data)
#cls_meta_lbl = cls_meta.label_info_list
#classifier = frame_meta.base_meta.batch_meta.classifier_meta_pool
#print('frame_meta:',classifier)
# label_info_list = pyds.NvDsLabelInfo.cast(classifier.label_info_list.data)
#print('label_info_list.result_class_id: {}'.format(label_info_list.result_class_id ))
#print('cls_meta.num_labels:',cls_meta.num_labels)
# print('cls_meta_lbl')
#cls_meta_lbl_info=pyds.NvDsLabelInfo.cast(cls_meta_lbl.data)
#result_str = str(cls_meta_lbl_info.result_label)#.tobytes().decode('iso-8859-1'))
#print("result_decode:", result_str)
except Exception as e: #StopIteration
break
#print(help(frame_meta))
#No primeiro frame pega altura e largura e atualiza o tempo da ultima deteccao para nao entrar em stop line
if self.frame_width == 0:
self.frame_width = frame_meta.source_frame_width
self.frame_height = frame_meta.source_frame_height
self.last_detection_moviment_timestamp = datetime.now()
self.time_continuous_detection = 0
print(frame_meta.source_frame_width, frame_meta.source_frame_height)
#Display meta para adicionar textos na imagem
display_meta=pyds.nvds_acquire_display_meta_from_pool(batch_meta)
# Me remova
self.cont_frame += 1
#Caso esteja em parada de linha, adicionar mensagem na imagem
if self.stop_line_status:
#Mensagem na imagem
self.add_text_to_display_meta(display_meta, "PARADA DE LINHA", int(self.frame_width/2), int(self.frame_height/2),1,1, font_size=40)
else:
#Adicionando no frame linha central que serve para contagem, isso é só para visualização
self.add_lines_to_display_meta(display_meta, self.line_count_x, 0, self.line_count_x, self.line_count_y)
#Objeto que contem as deteccoes
l_obj = frame_meta.obj_meta_list
#print(l_obj.class_id)
print("l_obj:",l_obj)
print("frame_meta.num_obj_meta:",frame_meta.num_obj_meta)
print("frame_meta.bInferDone:",frame_meta.bInferDone)
obj_meta=pyds.NvDsObjectMeta.cast(l_obj.data)
print("obj_meta:",obj_meta)
cls_obj = obj_meta.classifier_meta_list
print("cls_obj:",cls_obj)
Even when I use a yolov8m.onnx detector followed by the previously mentioned classifier, the result of classifier_meta_list still returns None.
I was able to obtain the classification result. However, since the classifier is secondary and requires tracking to be active, the detected object is only classified once.
Is there a parameter that enables classification on every frame for that object?
Thank you for reply. I developed a cascaded model where the detection is first performed using YOLOv8, and then the classification of the detected object is carried out with ResNet18. However, after the initial classification, since the object is static and maintains the same tracking ID, the classification is not repeated.
In my application, it is necessary to detect a static region where the classification should change based on the lighting condition: when the light is off, the region should belong to one class, and when the light is on, it should belong to another class. Because the region is static, the object’s ID does not change, preventing reclassification. For my application, it is crucial that this classification is performed continuously, even if the object remains the same. Is there a way to enable constant reclassification of the detected object?
Another question I have is how to apply the definition of an ROI (Region of Interest). Given that it’s a static region, instead of using a detector and a classifier in a cascade, would it be possible to simply define an ROI where the classification model could continuously perform the classification? How can I configure this ROI in the DeepStream pipeline?
The first two methods were unsuccessful. It was not possible to remove the tracker because, without it, the secondary model is unable to perform inference. Additionally, varying the values of matchingScoreWeight4VisualSimilarity did not yield good results, likely because it involves detecting a single, static region. The last suggested method will still be tested.
There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks
